计算机科学与技术（参考文献）A Context Aware Energy-Saving Scheme for Smart Camera Phones based on Activity Sensing

A Context Aware Energy-Saving Scheme for Smart Camera Phones based on Activity Sensing Yuanyuan Fan,Lei Xie,Yafeng Yin,Sanglu Lu State Key Laboratory for Novel Software Technology,Nanjing University,P.R.China Email:fyymonica@dislab.nju.edu.cn,Ixie@nju.edu.cn,yyf@dislab.nju.edu.cn,sanglu@nju.edu.cn Abstract-Nowadays more and more users tend to take photos time of smart camera phones.However,the previous work in with their smart phones.However,energy-saving continues to be energy-saving schemes for smart phones have the following a thorny problem for smart camera phones,since smart phone limitations:First,they mainly reduce the energy consumption photographing is a very power hungry function.In this paper, in a fairly isolated approach,without sufficiently considering we propose a context aware energy-saving scheme for smart the user's actual behaviors from the application perspective, camera phones,by accurately sensing the user's activities in the photographing process.Our solution is based on the observation this may greatly impact the user's experience of the smart that during the process of photographing,most of the energy are phones.Second,in regard to the energy-saving scheme for wasted in the preparations before the shooting.By leveraging photographing,they mainly focus on the shooting process the embedded sensors like the accelerometer and gyroscope,our instead of the preparations before shooting. solution is able to extract representative features to perceive the user's current activities including body movement,arm movement In this paper,we propose a context aware energy-saving and wrist movement.Furthermore,by maintaining an activity scheme for smart camera phones,by accurately sensing the state machine,our solution can accurately determine the user's user's activities in the photographing process.Our idea is current activity states and make the corresponding energy saving that,since current smart phones are mostly equipped with strategies.Experiment results show that,our solution is able to tiny sensors such as the accelerometer and gyroscope,we can perceive the user's activities with an average accuracy of 95.5% leverage these tiny sensors to effectively perceive the user's and reduce the overall energy consumption by 46.5%for smart activities,such that the corresponding energy-saving strategies camera phones compared to that without energy-saving scheme. can be applied according to the user's activities. There are several challenges in building an activity sensing- I.INTRODUCTION based scheme for smart phones.The first challenge is to ef- Nowadays smart phones have been widely used in our fectively classify the user's activities during the photographing daily lives.These devices are usually equipped with sensors process,which contains various levels of activities of bodies, such as the camera,accelerometer,and gyroscope.Due to arms and wrists.To address this challenge,we propose a the portability of smart phones,more and more people tend three-tier architecture for activity sensing,including the body to take photos with their smart phones.However,energy movement,arm movement and wrist movement.Furthermore, saving continues to be a upsetting problem for smart camera by maintaining an activity state machine,we can accurately phones,since smart phone photographing is a very power determine the user's current activity states and make the hungry function.For example,according to KS Mobile's [1] corresponding energy saving strategies.The second challenge report in 2014,the application Camera 360 Ultimate is listed is to make an appropriate trade-off between the accuracy of in the first place of top 10 battery draining applications for activity sensing and energy consumption.In order to accurately Android.Therefore,the huge energy consumption becomes a perceive the user's activities with the embedded sensors,more non-negligible pain point for the users of smart camera phones. types of sensor data and higher sampling rates are required. However,this further causes more energy consumption.To Nevertheless,during the process of photographing,we address this challenge,our solution only leverages those observe that a fairly large proportion of the energy is wasted low power sensors,such as accelerometer and gyroscope,to in the preparations before shooting.For example,the user first classify the activities by extracting representative features to turns on the camera in the smart phone,then the user may distinguish the user's activities respectively.We further choose move and adjust the camera phone time and again,so as to sampling rates according to the user's current activities.In this find a view,finally,the user focuses on the object and presses way,we can sufficiently reduce the energy consumption of the button to shoot.A lot of energy is wasted in the process activity sensing so as to achieve the overall energy efficiency. between two consecutive shootings,since the camera phone uses the same settings like the frame rate during the whole We make the following contributions in three folds:First, process,and these settings requires basically equal power we propose a context aware energy-saving scheme for smart consumption in the camera phone.Besides,it is also not wise camera phones,by leveraging the embedded sensor to conduct to frequently turn on/off the camera function,since it is not activity sensing.Based on the activity sensing results,we can only very annoying but also not energy efficient.Therefore, make the corresponding energy saving strategies.Second,we it is essential to reduce the unnecessary energy consumption build a three-tier architecture for activity sensing,including during the photographing process to greatly extend the life the body movement,arm movement and wrist movement.We use low-power sensors like the accelerometer and gyroscope Corresponding Author:Dr.Lei Xie,Ixie@nju.edu.cn to extract representative features to distinguish the user's

A Context Aware Energy-Saving Scheme for Smart Camera Phones based on Activity Sensing Yuanyuan Fan, Lei Xie, Yafeng Yin, Sanglu Lu State Key Laboratory for Novel Software Technology, Nanjing University, P.R. China Email: fyymonica@dislab.nju.edu.cn, lxie@nju.edu.cn, yyf@dislab.nju.edu.cn, sanglu@nju.edu.cn Abstract—Nowadays more and more users tend to take photos with their smart phones. However, energy-saving continues to be a thorny problem for smart camera phones, since smart phone photographing is a very power hungry function. In this paper, we propose a context aware energy-saving scheme for smart camera phones, by accurately sensing the user’s activities in the photographing process. Our solution is based on the observation that during the process of photographing, most of the energy are wasted in the preparations before the shooting. By leveraging the embedded sensors like the accelerometer and gyroscope, our solution is able to extract representative features to perceive the user’s current activities including body movement, arm movement and wrist movement. Furthermore, by maintaining an activity state machine, our solution can accurately determine the user’s current activity states and make the corresponding energy saving strategies. Experiment results show that, our solution is able to perceive the user’s activities with an average accuracy of 95.5% and reduce the overall energy consumption by 46.5% for smart camera phones compared to that without energy-saving scheme. I. INTRODUCTION Nowadays smart phones have been widely used in our daily lives. These devices are usually equipped with sensors such as the camera, accelerometer, and gyroscope. Due to the portability of smart phones, more and more people tend to take photos with their smart phones. However, energy￾saving continues to be a upsetting problem for smart camera phones, since smart phone photographing is a very power hungry function. For example, according to KS Mobile’s [1] report in 2014, the application Camera 360 Ultimate is listed in the first place of top 10 battery draining applications for Android. Therefore, the huge energy consumption becomes a non-negligible pain point for the users of smart camera phones. Nevertheless, during the process of photographing, we observe that a fairly large proportion of the energy is wasted in the preparations before shooting. For example, the user first turns on the camera in the smart phone, then the user may move and adjust the camera phone time and again, so as to find a view, finally, the user focuses on the object and presses the button to shoot. A lot of energy is wasted in the process between two consecutive shootings, since the camera phone uses the same settings like the frame rate during the whole process, and these settings requires basically equal power consumption in the camera phone. Besides, it is also not wise to frequently turn on/off the camera function, since it is not only very annoying but also not energy efficient. Therefore, it is essential to reduce the unnecessary energy consumption during the photographing process to greatly extend the life Corresponding Author: Dr. Lei Xie, lxie@nju.edu.cn time of smart camera phones. However, the previous work in energy-saving schemes for smart phones have the following limitations: First, they mainly reduce the energy consumption in a fairly isolated approach, without sufficiently considering the user’s actual behaviors from the application perspective, this may greatly impact the user’s experience of the smart phones. Second, in regard to the energy-saving scheme for photographing, they mainly focus on the shooting process instead of the preparations before shooting. In this paper, we propose a context aware energy-saving scheme for smart camera phones, by accurately sensing the user’s activities in the photographing process. Our idea is that, since current smart phones are mostly equipped with tiny sensors such as the accelerometer and gyroscope, we can leverage these tiny sensors to effectively perceive the user’s activities, such that the corresponding energy-saving strategies can be applied according to the user’s activities. There are several challenges in building an activity sensing￾based scheme for smart phones. The first challenge is to ef￾fectively classify the user’s activities during the photographing process, which contains various levels of activities of bodies, arms and wrists. To address this challenge, we propose a three-tier architecture for activity sensing, including the body movement, arm movement and wrist movement. Furthermore, by maintaining an activity state machine, we can accurately determine the user’s current activity states and make the corresponding energy saving strategies. The second challenge is to make an appropriate trade-off between the accuracy of activity sensing and energy consumption. In order to accurately perceive the user’s activities with the embedded sensors, more types of sensor data and higher sampling rates are required. However, this further causes more energy consumption. To address this challenge, our solution only leverages those low power sensors, such as accelerometer and gyroscope, to classify the activities by extracting representative features to distinguish the user’s activities respectively. We further choose sampling rates according to the user’s current activities. In this way, we can sufficiently reduce the energy consumption of activity sensing so as to achieve the overall energy efficiency. We make the following contributions in three folds: First, we propose a context aware energy-saving scheme for smart camera phones, by leveraging the embedded sensor to conduct activity sensing. Based on the activity sensing results, we can make the corresponding energy saving strategies. Second, we build a three-tier architecture for activity sensing, including the body movement, arm movement and wrist movement. We use low-power sensors like the accelerometer and gyroscope to extract representative features to distinguish the user’s

Walk L推UpAm Phone Rotate Fine-tuning and Shooting Lay Down Arm 40 80 100 20 40 Time(0.1s) Fig.1.The Process of Photographing activities.By maintaining an activity state machine,we can classify the user's activities in a very accurate approach.Third, we have implemented a system prototype in android-powered E10 smart camera phones,the experiment results shows that our solution is able to perceive the user's activities with an average accuracy of 95.5%and reduce the overall energy consumption by 46.5%for smart camera phones compared to that without energy-saving scheme. acc la gra gyro mf pro cam (a)Energy Consumption Compo-(b)Energy Consumption of Built-in nents Sensors II.SYSTEM OVERVIEW Fig.2.Energy Consumption.(a)pl-Samsung GT-19250,p2-Huawei H30- In order to adaptively reduce the power consumption based T10.p3-Samsung GT-19023,p4-Huawei G520-T10:(b)acc-accelerometer,la- on activity sensing,we first use the built-in sensors of the linear accelerometer,gra-gravity sensor.gyro-gyroscope,mf-magnetic field, phone to observe the human activities and discuss the energy pro-proxity,cam-camera. consumption during the process of photographing.Then,we introduce the system architecture of our proposed energy- preview mode.For the four phones,we can find that keeping saving scheme. screen on increases the power consumption by 20%,18%,27% and 26%,respectively.While using camera increases power A.Human activities related to photographing consumption by 63%,70%,53%and 57%,respectively.There- Usually,the users tend to have similar activities during pho- fore,in preparation for photographing,energy consumption is tographing.As shown in Fig.1,we use the linear accelerometer wasted on camera and screen. to detect the activities.Before or after the user takes photos. 2)Energy consumption of turning on/off the camera: he/she may stay motionless,walking or jogging.While taking Frequently turning on/off the phone is annoying and energy photos,the user usually lifts up the arm,rotates the phone, wasting.When we need to take photos frequently for a period makes fine-tuning,shoots a picture,then lays down the arm. of time,the camera may switch between on/off frequently and We categorize all the eight actions into the following three so is to screen.And user needs to press button and unlock parts: the screen with scrolling which results in large energy waste [3].Based on Table I and Eq.(1),we can find that energy Body movement:Motionless,walking and jogging TABLE I. ENERGY DATA OF TURNING ON/OFF PHONE (SAMSUNG Arm movement:Lifting up the arm and laying down GT-I9250) the arm Subjec Wrist movement:Rotating the phone,making fine- Energy Tumn On (Total) Tum Off (Total) Preview (1s) tuning and shooting a picture Energy(uAh】 769.71 400.71 16022 B.Energy consumption related to photographing consumption of turning on/off the phone one time can keep camera working in previewing mode for seven seconds,which Before we propose the energy-saving scheme to reduce the means one picture can be shot. power consumption based on the user's activities,we observe the energy consumption of the phone by using Monsoon power 769.71uAh+400.71uAh)/160.22Ah=7.305s(1) monitor [2]firstly It indicates that frequently turning on/off the phone manually 1)Energy consumption in preparation for photographing: is indeed inconvenient and rather energy-consuming. Power consumption for shooting a picture is large.We observe 3)Energy consumption of the sensors:Fig.2(b)shows the the power consumption in the following four android-based power consumption of a phone when a sensor is turned on. phones,i.e.,Samsung GT-19250,Huawei H30-T10,Samsung All these sensors work in their maximum sampling frequency, GT-19230 and Huawei G520-T10.In Fig.2(a),"Base"repre- i.e.,100Hz.When we turn on the camera and phone will be sents the phone's basic power when the phone is off."Display" in preview state,its increase of power is much larger than that represents the power when phone is idle and only keeps screen of other sensors.Therefore,low power-consuming sensors can on."Camera"represents the power when camera works in be used to reduce the energy consumption of photographing

Time (0.1s) 0 20 40 60 80 100 120 140 150 Linear Accelerometer -6 -4 -2 0 2 4 6 8 x-axis Walk Lift Up Arm Phone Rotate Fine-tuning and Shooting Lay Down Arm Walk A B C D Fig. 1. The Process of Photographing activities. By maintaining an activity state machine, we can classify the user’s activities in a very accurate approach. Third, we have implemented a system prototype in android-powered smart camera phones, the experiment results shows that our solution is able to perceive the user’s activities with an average accuracy of 95.5% and reduce the overall energy consumption by 46.5% for smart camera phones compared to that without energy-saving scheme. II. SYSTEM OVERVIEW In order to adaptively reduce the power consumption based on activity sensing, we first use the built-in sensors of the phone to observe the human activities and discuss the energy consumption during the process of photographing. Then, we introduce the system architecture of our proposed energy￾saving scheme. A. Human activities related to photographing Usually, the users tend to have similar activities during pho￾tographing. As shown in Fig. 1, we use the linear accelerometer to detect the activities. Before or after the user takes photos, he/she may stay motionless, walking or jogging. While taking photos, the user usually lifts up the arm, rotates the phone, makes fine-tuning, shoots a picture, then lays down the arm. We categorize all the eight actions into the following three parts: • Body movement: Motionless, walking and jogging • Arm movement: Lifting up the arm and laying down the arm • Wrist movement: Rotating the phone, making fine￾tuning and shooting a picture B. Energy consumption related to photographing Before we propose the energy-saving scheme to reduce the power consumption based on the user’s activities, we observe the energy consumption of the phone by using Monsoon power monitor [2] firstly. 1) Energy consumption in preparation for photographing: Power consumption for shooting a picture is large. We observe the power consumption in the following four android-based phones, i.e., Samsung GT-I9250, Huawei H30-T10, Samsung GT-I9230 and Huawei G520-T10. In Fig. 2(a), “Base” repre￾sents the phone’s basic power when the phone is off. “Display” represents the power when phone is idle and only keeps screen on. “Camera” represents the power when camera works in p1 p2 p3 p4 Power (mW) 0 500 1000 1500 2000 2500 Base Display Camera (a) Energy Consumption Compo￾nents acc la gra gyro mf pro cam Power (mW) 0 500 1000 1500 (b) Energy Consumption of Built-in Sensors Fig. 2. Energy Consumption. (a) p1-Samsung GT-I9250, p2-Huawei H30- T10, p3-Samsung GT-I9023, p4-Huawei G520-T10; (b) acc-accelerometer, la￾linear accelerometer, gra-gravity sensor, gyro-gyroscope, mf-magnetic field, pro-proxity, cam-camera. preview mode. For the four phones, we can find that keeping screen on increases the power consumption by 20%, 18%, 27% and 26%, respectively. While using camera increases power consumption by 63%, 70%, 53% and 57%, respectively. There￾fore, in preparation for photographing, energy consumption is wasted on camera and screen. 2) Energy consumption of turning on/off the camera: Frequently turning on/off the phone is annoying and energy wasting. When we need to take photos frequently for a period of time, the camera may switch between on/off frequently and so is to screen. And user needs to press button and unlock the screen with scrolling which results in large energy waste [3]. Based on Table I and Eq. (1), we can find that energy TABLE I. ENERGY DATA OF TURNING ON/OFF PHONE (SAMSUNG GT-I9250) Energy Subject Turn On (Total) Turn Off (Total) Preview (1s) Energy (uAh) 769.71 400.71 160.22 consumption of turning on/off the phone one time can keep camera working in previewing mode for seven seconds, which means one picture can be shot. (769.71uAh + 400.71uAh)/160.22uAh = 7.305s (1) It indicates that frequently turning on/off the phone manually is indeed inconvenient and rather energy-consuming. 3) Energy consumption of the sensors: Fig. 2(b) shows the power consumption of a phone when a sensor is turned on. All these sensors work in their maximum sampling frequency, i.e., 100Hz. When we turn on the camera and phone will be in preview state, its increase of power is much larger than that of other sensors. Therefore, low power-consuming sensors can be used to reduce the energy consumption of photographing

Sensors -Sensor Data- Segmentation -Data Segment (Choosing Low-Power Sensors) (Based on Pause between Actions) Linear Accelerometer Gyroscope Activity Sensing (Recognizing the activities using state machine) Gravity Sensor Set A Set B -State- Motionless Arm Lifting Up Mobile Rotating Energy Saving (Applying Different Schemes Walking Fine-Tuning pased on States) Maximum Body Energy Saved Jogging Arm Laying Down Shooting Medium Energy Saved Arm Body Arm Wrist Minimum Wrist Energy Saved Fig.3.System Architecture L00 According to the above observations,we can utilize the low energy-consuming built-in sensors of the phone to detect z-axis top the user's activities and reduce the energy consumption of taking photos.A simple example could be turning off the screen,decreasing the brightness of the screen,or decreasing x-axis the preview frame rate of the camera to reduce the energy cost when we find the user is not taking a photo. (a)Coordinates of Mo-(b)hold horizontally (c)hold horizontally C.System Architecture tion Sensors naturally backward The architecture of our system is shown in Figure 3. Fig.4.Coordinates of Phone and Direction of Phone Hold Firstly,we mainly obtain the data from low power-consuming built-in sensors,i.e..the linear accelerometer,the gyroscope and the gravity sensor,as shown in the "Sensor"module. Arm level:It includes lifting the arm up and laying Secondly.we separate the data into different regions,which the arm down.The relationship between the data of are corresponding to the users'actions,as shown in the gravity sensor and linear accelerometer is used to "Activity Sensing"module.Thirdly,we adaptively adopt an distinguish the two actions.And voting mechanism appropriate energy-saving scheme for each action,as shown is used to guarantee the accuracy. in the "Energy Saving"module.In the following paragraphs, we briefly introduce how we can realize the activity sensing Wrist level:It includes rotating the phone,making and reduce the power consumption. fine-tuning,and shooting a picture.We make use of 1)Activity Sensing:Based on section II-A,the user's ac- a linear SVM model to distinguish them with the variance,mean,maximum and minimum of three tions can be categorized into three parts:body movement,arm movement,wrist movement.Correspondingly,in our system axises of three sensors as its features. architecture,we call the above parts as body level,arm level, wrist level,respectively.In each level,there may be more than 2)Energy-saving Scheme:Based on the feature and energy one action.Besides,the different levels may exist some transfer consumption in each action/state,we propose an adaptive relations.Therefore,we use the State Machine to describe the energy-saving scheme for taking photos.For example,when specific actions of the user.In the State Machine,each action you walk,jog or stay motionless,it's unnecessary to keep the is represented as a state.The transferable relations between screen on.When you lift your arm up,it's better to turn on the states are shown in Fig.3.Before we determine the type the screen and adjust the screen's brightness based on the light of the action,we first estimate which level the action belongs conditions.When you make fine-tuning to observe the camera to.Then,we further infer the specific action of the user based view before shooting a picture,it's better to make the camera on more sensor information. work on the preview state.In this way,we can make the context aware energy-saving schemes for the camera phones. Body level:It includes motionless,walking and jog- ging.Motionlessness can be recognized with its low variance of linear accelerometer's data.Then walk- III.SYSTEM DESIGN ing and jogging are distinguished with the frequency which can be calculated using Fast Fourier Transfor- In this section,we present the design of our energy-saving mation. scheme for smart camera phones based on activity sensing

Sensors Linear Accelerometer Gyroscope Gravity Sensor Segmentation Activity Sensing Body Arm Wrist Arm Lifting Up Arm Laying Down Mobile Rotating Fine-Tuning Shooting Motionless Walking Jogging Set A Set B Maximum Energy Saved Body Medium Energy Saved Arm Minimum Energy Saved Wrist Energy Saving (Choosing Low-Power Sensors) (Based on Pause between Actions) Sensor Data Data Segment (Applying Different Schemes based on States) State (Recognizing the activities using state machine) Fig. 3. System Architecture According to the above observations, we can utilize the low energy-consuming built-in sensors of the phone to detect the user’s activities and reduce the energy consumption of taking photos. A simple example could be turning off the screen, decreasing the brightness of the screen, or decreasing the preview frame rate of the camera to reduce the energy cost when we find the user is not taking a photo. C. System Architecture The architecture of our system is shown in Figure 3. Firstly, we mainly obtain the data from low power-consuming built-in sensors, i.e., the linear accelerometer, the gyroscope and the gravity sensor, as shown in the “Sensor” module. Secondly, we separate the data into different regions, which are corresponding to the users’ actions, as shown in the “Activity Sensing” module. Thirdly, we adaptively adopt an appropriate energy-saving scheme for each action, as shown in the “Energy Saving” module. In the following paragraphs, we briefly introduce how we can realize the activity sensing and reduce the power consumption. 1) Activity Sensing: Based on section II-A, the user’s ac￾tions can be categorized into three parts: body movement, arm movement, wrist movement. Correspondingly, in our system architecture, we call the above parts as body level, arm level, wrist level, respectively. In each level, there may be more than one action. Besides, the different levels may exist some transfer relations. Therefore, we use the State Machine to describe the specific actions of the user. In the State Machine, each action is represented as a state. The transferable relations between the states are shown in Fig. 3. Before we determine the type of the action, we first estimate which level the action belongs to. Then, we further infer the specific action of the user based on more sensor information. • Body level: It includes motionless, walking and jog￾ging. Motionlessness can be recognized with its low variance of linear accelerometer’s data. Then walk￾ing and jogging are distinguished with the frequency which can be calculated using Fast Fourier Transfor￾mation. x-axis y-axis z-axis top (a) Coordinates of Mo￾tion Sensors top top (b) hold horizontally naturally top top (c) hold horizontally backward Fig. 4. Coordinates of Phone and Direction of Phone Hold • Arm level: It includes lifting the arm up and laying the arm down. The relationship between the data of gravity sensor and linear accelerometer is used to distinguish the two actions. And voting mechanism is used to guarantee the accuracy. • Wrist level: It includes rotating the phone, making fine-tuning, and shooting a picture. We make use of a linear SVM model to distinguish them with the variance, mean, maximum and minimum of three axises of three sensors as its features. 2) Energy-saving Scheme: Based on the feature and energy consumption in each action/state, we propose an adaptive energy-saving scheme for taking photos. For example, when you walk, jog or stay motionless, it’s unnecessary to keep the screen on. When you lift your arm up, it’s better to turn on the screen and adjust the screen’s brightness based on the light conditions. When you make fine-tuning to observe the camera view before shooting a picture, it’s better to make the camera work on the preview state. In this way, we can make the context aware energy-saving schemes for the camera phones. III. SYSTEM DESIGN In this section, we present the design of our energy-saving scheme for smart camera phones based on activity sensing

End of a 40 80 80 100 2 180 Time (0.1s) Fig.5.Segment the Data of Linear Accelerometer A.Activity Sensing i00 1)Raw Data Collection:We collect data from linear ac- otionlessness celerometer,gyroscope and gravity sensor of android phones. These sensors have their own coordinates as shown in Figures 4(a),which are different from the earth coordinate system 100120 180 For example,when we hold the phone as Fig.4(b),the data 0 Variance of gravity sensor's x-axis almost equals to g(9.8m/s2).When we hold the phone as Fig.4(c),the result is opposite. Fig.7. CDF of Variance of Y-axis of Three Actions in Body Level 2)Action Segmentation:From sensors,we can only get sequential raw data.To do the following activity sensing,data point of the next segment and return back to calculate the should be segmented as one segment corresponds to one action. variance in the window of linear accelerometer's data.Fourth Observation.For a user,there is always a short pause if there is too much data before the second eligible variance between two different actions shown with red rectangles (A showing up,a maximum segment size is set to segment data. B,and D)in Figure 1.However,some actions like fine- The maximum size is set as ten times of the value of sensor's tuning and shooting are very gentle,it's difficult for the linear sampling rate because most of the actions in arm and wrist accelerometer to detect the pause from the actions shown with level won't last for more than 10 seconds for common people. blue rectangle (D)in Figure 5.On this occasion,gyroscope is 3)Action Recognition:We first do the recognition in three used to assist for the segmentation because it's more sensitive levels respectively.Then we describe how to do recognition with the motion.The gyroscope data corresponding to the data among levels. in rectangle D is shown in Figures 6 and the pause between actions can be detected as shown with red rectangles (D1/D2). Body Level:Body level includes three actions which Back to Figure 5,one action which lasts for a long time shown are motionlessness,walking and jogging.They are important with purple rectangle(E),may bring computational overhead. actions which connect two shoots.As the movements of walking and jogging are very obvious,we take advantage of linear accelerometer to classify the actions. 0 Observation.Motionlessness is easy to be distinguished D1 D2 x-axis from walk and jog because of its low variance of raw data from linear accelerometer.Figures 7 shows the distribution of three z-axis actions'variances.Variance of motionlessness is almost zero 80 85 90 95 100 105 110 Time (0.1s) and can be clearly distinguished.While walking and jogging can't be distinguished only based on the variances as they have Fig.6.Raw Data of Gyroscope Corresponding to Rectangle D in Fig 5 some overlaps. Segmentation.First,we leverage a sliding window to We hold the phone like Fig 4(b)and use linear accelerom- continuously calculate the variances of data of linear ac- eter to get raw data of walking and jogging shown in Figures celerometer's three axises.Second,if all three variances are 8(a)and 8(c).We apply Fast Fourier Transform on the data and below a threshold,the window is regraded as the start/end of get the spectrum.Fig 8(b)shows that the frequency of walking a segment shown with green rectangle (B/C)in Fig 5.The is about 1 Hz.Fig 8(d)shows that the frequency of jogging window size is set as half of the value of sensor's sampling is about 3.5 Hz.Thus,these two actions can be distinguished frequency as the pause between two continuous actions is using frequency always less than half a second.Third,when two continuous Recognition in Body Level.Effected by the holding gesture, sliding windows whose variances are both below the threshold, the changes of data in three axises are different.(1)We first we use the corresponding data of gyroscope and calculate the determine the axis whose data will be used.To common variance in the sliding window.If two continuous windows people,the phone can be held perpendicularly or parallel whose variances are below the threshold in gyroscope,we to the ground.If the phone is held perpendicularly to the continue to calculate until a window whose variance above ground,the data of z-axis doesn't change a lot in this level.If threshold is found.Then,this part is regarded as a segment. phone is held parallel to the ground,the data of x-axis doesn't After that,we will take last sample of the window as a start change obviously.Therefore,we use the data of y-axis.(2)We

Time (0.1s) 0 20 40 60 80 100 120 140 160 180 Linear Accelerometer -10 -5 0 5 10 x-axis y-axis z-axis Start of a Segment End of a Segment A B C D E Sliding Window Fig. 5. Segment the Data of Linear Accelerometer A. Activity Sensing 1) Raw Data Collection: We collect data from linear ac￾celerometer, gyroscope and gravity sensor of android phones. These sensors have their own coordinates as shown in Figures 4(a), which are different from the earth coordinate system. For example, when we hold the phone as Fig. 4(b), the data of gravity sensor’s x-axis almost equals to g (9.8m/s2 ). When we hold the phone as Fig. 4(c) , the result is opposite. 2) Action Segmentation: From sensors, we can only get sequential raw data. To do the following activity sensing, data should be segmented as one segment corresponds to one action. Observation. For a user, there is always a short pause between two different actions shown with red rectangles (A, B, and D) in Figure 1. However, some actions like fine￾tuning and shooting are very gentle, it’s difficult for the linear accelerometer to detect the pause from the actions shown with blue rectangle (D) in Figure 5. On this occasion, gyroscope is used to assist for the segmentation because it’s more sensitive with the motion. The gyroscope data corresponding to the data in rectangle D is shown in Figures 6 and the pause between actions can be detected as shown with red rectangles (D1/D2). Back to Figure 5, one action which lasts for a long time shown with purple rectangle (E), may bring computational overhead. Time (0.1s) 80 85 90 95 100 105 110 Gyroscope -4 -2 0 2 x-axis y-axis z-axis D1 D2 Fig. 6. Raw Data of Gyroscope Corresponding to Rectangle D in Fig 5 Segmentation. First, we leverage a sliding window to continuously calculate the variances of data of linear ac￾celerometer’s three axises. Second, if all three variances are below a threshold, the window is regraded as the start/end of a segment shown with green rectangle (B/C) in Fig 5. The window size is set as half of the value of sensor’s sampling frequency as the pause between two continuous actions is always less than half a second. Third, when two continuous sliding windows whose variances are both below the threshold, we use the corresponding data of gyroscope and calculate the variance in the sliding window. If two continuous windows whose variances are below the threshold in gyroscope, we continue to calculate until a window whose variance above threshold is found. Then, this part is regarded as a segment. After that, we will take last sample of the window as a start Variance 0 20 40 60 80 100 120 140 160 180 200 CDF 0 0.2 0.4 0.6 0.8 1 jog walk motionlessness 5 Fig. 7. CDF of Variance of Y-axis of Three Actions in Body Level point of the next segment and return back to calculate the variance in the window of linear accelerometer’s data. Fourth, if there is too much data before the second eligible variance showing up, a maximum segment size is set to segment data. The maximum size is set as ten times of the value of sensor’s sampling rate because most of the actions in arm and wrist level won’t last for more than 10 seconds for common people. 3) Action Recognition: We first do the recognition in three levels respectively. Then we describe how to do recognition among levels. Body Level: Body level includes three actions which are motionlessness, walking and jogging. They are important actions which connect two shoots. As the movements of walking and jogging are very obvious, we take advantage of linear accelerometer to classify the actions. Observation. Motionlessness is easy to be distinguished from walk and jog because of its low variance of raw data from linear accelerometer. Figures 7 shows the distribution of three actions’ variances. Variance of motionlessness is almost zero and can be clearly distinguished. While walking and jogging can’t be distinguished only based on the variances as they have some overlaps. We hold the phone like Fig 4(b) and use linear accelerom￾eter to get raw data of walking and jogging shown in Figures 8(a) and 8(c). We apply Fast Fourier Transform on the data and get the spectrum. Fig 8(b) shows that the frequency of walking is about 1 Hz. Fig 8(d) shows that the frequency of jogging is about 3.5 Hz. Thus, these two actions can be distinguished using frequency. Recognition in Body Level. Effected by the holding gesture, the changes of data in three axises are different. (1) We first determine the axis whose data will be used. To common people, the phone can be held perpendicularly or parallel to the ground. If the phone is held perpendicularly to the ground, the data of z-axis doesn’t change a lot in this level. If phone is held parallel to the ground, the data of x-axis doesn’t change obviously. Therefore, we use the data of y-axis. (2) We

三 0 10 (a):Lift Up with Phone Rotating 0 2 (a):Time (0.1s) (b):Frequency (Hz) (b):Lift Up with Phone Rotating 0 -e-Product 0 (c):Lift Up with Phone Rotating (c):Time (0.1s) (d):Frequency (Hz) Fig.10.Lift up the phone with rotating 360 degrees Fig.8.Frequencies of walking and jogging.(a)shows raw data of walking and (b)shows its frequency.(c)shows raw data of jogging and (d)shows its direction.Therefore only the data of linear accelerometer's x- frequency. axis is showed.In Figures 9(a),when lifting up your arm,the value of gravity sensor stays positive while the value of linear 10 10 Gravity x-axis x-axis accelerometer changes from positive to negative.It means the -Linear Acc x-axis Linear Acc x-axis signs of two sensors'value change from same to different. 4 64 In Figures 9(b),the signs of two sensors'value change from 2 2 different to same.When the phone is held as Figures 4(c).the 0 corresponding sensor data is showed in Figures 9(c)and (d) When lifting up your arm,the signs of x-axis value of two (a):Lift up arm when phone (b):Lay down arm when phone sensors still change form same to different.And when laying hold horizontal normal hold horizontal normal down your arm,the result is opposite. However,the phone may be held in hand in any gesture. 0 We lift up the phone and rotate 360 degrees at the same time. 2 The data of gravity sensor and linear accelerometer is showed in Figures 10 (a)and (b).We can't simply figure out the Gravity x-axis Gravity x-axis Linear Acc x-axis 的 -Linear Acc x-axis relationship between two sensors.The data in three axises of 10 gravity sensor,whose absolute value is maximum,is chosen (c):Lift up arm when phone (d):Lay down arm when phone as shown with black circle in Fig 10(a).And the data of linear hold horizontal backward hold horizontal backward accelerometer of corresponding axis is chosen,as shown with black circles in Fig 10(b).We multiply the two corresponding Fig.9.Data of linear acceleration and gravity sensor of arm level when data and the result is showed in Figures 10(c).We can find phone held horizontally in normal and in backward direction that the sign changes from positive to negative.In summary, when you lift arm up,the signs of gravity sensor and linear calculate the variance of y-axis of linear accelerometer.If it is accelerometer change from same to different.When you lay less than a threshold,the action is regarded as motionlessness. down arm,the change is diametrically opposite. The threshold is set to 5 according to Figure 7.(3)If the action Recognition in Arm Level.The maximum absolute data of is not motionlessness,we apply FFT to the data segment.In gravity sensor and the corresponding data of linear accelerom- general,the frequency of walk ranges from I Hz to 3 Hz and eter are chosen.Then we analyze the relationship between that of jog is 3 Hz to 6 Hz.Therefore,if the frequency is the two sensors and then apply voting mechanism to avoid less than 3.the action is recognized as walking.Otherwise. the noise made by hands tremble.At last,if the signs of two it's jogging. sensors'selected data change from same to different,the state Arm Level:Arm level contains two actions,arm lifting up should be arm lifting up.Otherwise,the state should be arm laying down.The specific process is showed in Algorithm 1. and laying down.They are actions which connect body and wrist level.After you lift up your arm,you may start shooting Wrist Level:Wrist movement contains phone rotating, After you lay down your arm,the shooting may end. fine-tuning and shooting.Picture is shot in this level Observation.Arms lifting up and laying down are two Observation.The raw data of three axises of linear ac- reversed actions.When we hold the phone horizontally in celerometer of three actions are showed in Figures 11(a)and normal direction as Fig 4(b),we get sensors'data of lifting arm 11(b).From Figure 11(a),phone rotating can be distinguished up shown in Fig 9(a)and laying arm down shown in Figures by using a plane.From Figure 11(b),the other two actions 9 (b).Under this situation,the data of linear accelerometer can be distinguished.Therefore,a classifier as Support Vector will change mostly in x-axis as the motion happens in its Machine (SVM)can be used for classification.We take the

(a): Time (0.1s) 0 10 20 30 40 50 Linear Acceleration (m/s 2 ) -6 -4 -2 0 2 4 6 8 x-axis y-axis z-axis (b): Frequency (Hz) 012345 Amplitude 0 20 40 60 80 100 x-axis y-axis z-axis (c): Time (0.1s) 0 20 40 60 Linear Acceleration (m/s 2 ) -15 -10 -5 0 5 10 15 x-axis y-axis z-axis (d): Frequency (Hz) 012345 Amplitude 0 50 100 150 200 x-axis y-axis z-axis Fig. 8. Frequencies of walking and jogging. (a) shows raw data of walking and (b) shows its frequency. (c) shows raw data of jogging and (d) shows its frequency. (a): Lift up arm when phone hold horizontal normal 0 2 4 6 8 10 Sensor Data -4 -2 0 2 4 6 8 10 Gravity x-axis Linear Acc x-axis (b): Lay down arm when phone hold horizontal normal 0 2 4 6 8 10 Sensor Data -4 -2 0 2 4 6 8 10 Gravity x-axis Linear Acc x-axis (c): Lift up arm when phone hold horizontal backward 0 2 4 6 8 10 Sensor Data -10 -8 -6 -4 -2 0 2 4 Gravity x-axis Linear Acc x-axis (d): Lay down arm when phone hold horizontal backward 0 2 4 6 8 10 Sensor Data -10 -8 -6 -4 -2 0 2 Gravity x-axis Linear Acc x-axis Fig. 9. Data of linear acceleration and gravity sensor of arm level when phone held horizontally in normal and in backward direction calculate the variance of y-axis of linear accelerometer. If it is less than a threshold, the action is regarded as motionlessness. The threshold is set to 5 according to Figure 7. (3) If the action is not motionlessness, we apply FFT to the data segment. In general, the frequency of walk ranges from 1 Hz to 3 Hz and that of jog is 3 Hz to 6 Hz. Therefore, if the frequency is less than 3, the action is recognized as walking. Otherwise, it’s jogging. Arm Level: Arm level contains two actions, arm lifting up and laying down. They are actions which connect body and wrist level. After you lift up your arm, you may start shooting. After you lay down your arm, the shooting may end. Observation. Arms lifting up and laying down are two reversed actions. When we hold the phone horizontally in normal direction as Fig 4(b), we get sensors’ data of lifting arm up shown in Fig 9(a) and laying arm down shown in Figures 9 (b). Under this situation, the data of linear accelerometer will change mostly in x-axis as the motion happens in its (a): Lift Up with Phone Rotating Gravity Sensor 0 1 2 3 4 5 6 7 8 9 10 -10 -5 0 5 10 x-axis y-axis z-axis biggest absolute data (b): Lift Up with Phone Rotating 0 1 2 3 4 5 6 7 8 9 10 Linear Accelerometer -6 -4 -2 0 2 x-axis y-axis z-axis corresponding data (c): Lift Up with Phone Rotating 0 1 2 3 4 5 6 7 8 9 10 Product of Sensors -40 -20 0 20 40 60 Product Fig. 10. Lift up the phone with rotating 360 degrees direction. Therefore only the data of linear accelerometer’s x￾axis is showed. In Figures 9(a), when lifting up your arm, the value of gravity sensor stays positive while the value of linear accelerometer changes from positive to negative. It means the signs of two sensors’ value change from same to different. In Figures 9(b), the signs of two sensors’ value change from different to same. When the phone is held as Figures 4(c), the corresponding sensor data is showed in Figures 9(c) and (d). When lifting up your arm, the signs of x-axis value of two sensors still change form same to different. And when laying down your arm, the result is opposite. However, the phone may be held in hand in any gesture. We lift up the phone and rotate 360 degrees at the same time. The data of gravity sensor and linear accelerometer is showed in Figures 10 (a) and (b). We can’t simply figure out the relationship between two sensors. The data in three axises of gravity sensor, whose absolute value is maximum, is chosen as shown with black circle in Fig 10(a). And the data of linear accelerometer of corresponding axis is chosen, as shown with black circles in Fig 10(b). We multiply the two corresponding data and the result is showed in Figures 10(c). We can find that the sign changes from positive to negative. In summary, when you lift arm up, the signs of gravity sensor and linear accelerometer change from same to different. When you lay down arm, the change is diametrically opposite. Recognition in Arm Level. The maximum absolute data of gravity sensor and the corresponding data of linear accelerom￾eter are chosen. Then we analyze the relationship between the two sensors and then apply voting mechanism to avoid the noise made by hands tremble. At last, if the signs of two sensors’ selected data change from same to different, the state should be arm lifting up. Otherwise, the state should be arm laying down. The specific process is showed in Algorithm 1. Wrist Level: Wrist movement contains phone rotating, fine-tuning and shooting. Picture is shot in this level. Observation. The raw data of three axises of linear ac￾celerometer of three actions are showed in Figures 11(a) and 11(b). From Figure 11(a), phone rotating can be distinguished by using a plane. From Figure 11(b), the other two actions can be distinguished. Therefore, a classifier as Support Vector Machine (SVM) can be used for classification. We take the

Algorithm 1:Recognition in Arm Level celerometer,gyroscope and gravity sensor in the data segment Input:Data of three axises of gravity sensor and linear then take advantage of trained SVM model to predict the acceleromter current state Output:Arm state Recognition among Three Levels:We use an activity 1 Calculate the absolute data of all three axises of gravity state machine to help us do the activity recognition which sensor is shown in Figure 3.All the eight actions are eight states and 2 Get the axis position set A;where absolute data is the arrows show the relationship among the actions.The eight maximum among every three samples states belong to three levels,which are body,arm and wrist. 3 Get the original gravity data set Gi according to axis In each level,any two of the states can transform mutually. position set Ai,get the linear accelerometer data set LA;according to axis position set Ai However,the states in body level only can transform to arm 4 Multiply the corresponding data in set Gi and LA;and lifting up state.And only arm laying down state can transform store the sign of the result in set S to the state in body level.The connection of arm and wrist level s Voting the result in the first half of set Si and store the is similar.Based on the relationship,we divided all these states sign result as Sign,voting the result in the second into two sets (set A and B)shown in Figure 3.What's more, half and store the sign result as Sign2 once we press the button to take the picture,we calibrate the 6 if Sign is positive and Sign2 is negative then state if there's somthing wrong.Making use of the relationship 7 return Lifting arm up and the calibration,we do activity recognition as shown in Algorithm 2 s if Sign is negative and Sign2 is positive then return Laying arm down Algorithm 2:Recognition among Three Levels Input:Array A,which stores eight states Output:Current state Cs 1 Search A;and get the last state l 2 Recognize in arm level and assign the result to Arms 3 if Arms is in arm level then 4 Cs is assigned to Arms s else a 6 x-axis if the button of shotting is pressed then (a)Data of Three Actions Shown in (b)Data of Fine-tuning and Shooting C.is assigned to shooting X-Y Shown in X-Y-Z else if ls is in Set A then Fig.11.Data of Linear Accelerometer of Three Actions in Wrist Level 10 Recognize in body level and Cs is assigned to the result of recognition in body level else characteristics of linear accelerometer,gyroscope and gravity Recognize in wrist level and Cs is assigned sensor into account.Linear accelerometer can detect the abso to the result of recognition in wrist level lute motion.Gyroscope is sensitive to the rotation of phone Gravity sensor shows holding gesture.Therefore,using all 13 Update Ai these sensors'data to classify the actions can achieve good 14 return Cs effect.The motions and movement range of these actions are different.Therefore,we extract the variance,mean,maximum, minimum of three axises as features and use linear kernel to train a SVM model to classify three actions.We also try to train B.Energy-saving Scheme the SVM model with other six combinations of sensors and Based on the state obtained from activity sensing module three different kernels shown in Figure 12.And the accuracy and the feature of different action,an adaptive energy-saving of using three sensors with linear kernel is highest. scheme is proposed. 1)Energy Saving Points:Based on the analysis in section II-A,we have known the camera and screen are two energy- 0.8 hungry parts. Observation of Camera.To camera,the resolution of the photo and frame rate of preview are possible impact factors of energy consumption.However,high or low resolution has less 0 effect to energy consumption shown in Fig.13. RBF SVM Poly SVi Before shooting,the camera is in preview mode.The frame rate of preview has relationship with energy consumption Fig.12.Accuracies of Different SVM Models shown in Figures 14(a).We conduct the experiments with Samsung phone(GT-19250).The x-labels represent the ranges Recognition in Wrist Level.We first calculate the variance, of frame rate in preview mode.We discover that when the mean,maximum,minimum of all three axises of linear ac- range is 15-15,the energy consumption is minimum.Because

Algorithm 1: Recognition in Arm Level Input: Data of three axises of gravity sensor and linear acceleromter Output: Arm state 1 Calculate the absolute data of all three axises of gravity sensor 2 Get the axis position set Ai where absolute data is maximum among every three samples 3 Get the original gravity data set Gi according to axis position set Ai , get the linear accelerometer data set LAi according to axis position set Ai 4 Multiply the corresponding data in set Gi and LAi and store the sign of the result in set Si 5 Voting the result in the first half of set Si and store the sign result as Sign1, voting the result in the second half and store the sign result as Sign2 6 if Sign1 is positive and Sign2 is negative then 7 return Lifting arm up 8 if Sign1 is negative and Sign2 is positive then 9 return Laying arm down x-axis -2 -1.5 -1 -0.5 0 0.5 1 y-axis -2.5 -2 -1.5 -1 -0.5 0 0.5 1 fine-tuning shooting phone rotating plane (a) Data of Three Actions Shown in X-Y (b) Data of Fine-tuning and Shooting Shown in X-Y-Z Fig. 11. Data of Linear Accelerometer of Three Actions in Wrist Level characteristics of linear accelerometer, gyroscope and gravity sensor into account. Linear accelerometer can detect the abso￾lute motion. Gyroscope is sensitive to the rotation of phone. Gravity sensor shows holding gesture. Therefore, using all these sensors’ data to classify the actions can achieve good effect. The motions and movement range of these actions are different. Therefore, we extract the variance, mean, maximum, minimum of three axises as features and use linear kernel to train a SVM model to classify three actions. We also try to train the SVM model with other six combinations of sensors and three different kernels shown in Figure 12. And the accuracy of using three sensors with linear kernel is highest. Linear SVM RBF SVM Poly SVM Accuray 0 0.2 0.4 0.6 0.8 1 Linear Accelerometer Gyroscope Gravity Linear Acc. + Gravity Linear Acc. + Gyroscope Gravity + Gyroscope Linear Acc + Gravity + Gyroscope Fig. 12. Accuracies of Different SVM Models Recognition in Wrist Level. We first calculate the variance, mean, maximum, minimum of all three axises of linear ac￾celerometer, gyroscope and gravity sensor in the data segment then take advantage of trained SVM model to predict the current state. Recognition among Three Levels: We use an activity state machine to help us do the activity recognition which is shown in Figure 3. All the eight actions are eight states and the arrows show the relationship among the actions. The eight states belong to three levels, which are body, arm and wrist. In each level, any two of the states can transform mutually. However, the states in body level only can transform to arm lifting up state. And only arm laying down state can transform to the state in body level. The connection of arm and wrist level is similar. Based on the relationship, we divided all these states into two sets (set A and B) shown in Figure 3. What’s more, once we press the button to take the picture, we calibrate the state if there’s somthing wrong. Making use of the relationship and the calibration, we do activity recognition as shown in Algorithm 2. Algorithm 2: Recognition among Three Levels Input: Array Ai which stores eight states Output: Current state Cs 1 Search Ai and get the last state ls 2 Recognize in arm level and assign the result to Arms 3 if Arms is in arm level then 4 Cs is assigned to Arms 5 else 6 if the button of shotting is pressed then 7 Cs is assigned to shooting 8 else 9 if ls is in Set A then 10 Recognize in body level and Cs is assigned to the result of recognition in body level 11 else 12 Recognize in wrist level and Cs is assigned to the result of recognition in wrist level 13 Update Ai 14 return Cs B. Energy-saving Scheme Based on the state obtained from activity sensing module and the feature of different action, an adaptive energy-saving scheme is proposed. 1) Energy Saving Points: Based on the analysis in section II-A, we have known the camera and screen are two energy￾hungry parts. Observation of Camera. To camera, the resolution of the photo and frame rate of preview are possible impact factors of energy consumption. However, high or low resolution has less effect to energy consumption shown in Fig. 13. Before shooting, the camera is in preview mode. The frame rate of preview has relationship with energy consumption shown in Figures 14(a). We conduct the experiments with Samsung phone (GT-I9250). The x-labels represent the ranges of frame rate in preview mode. We discover that when the range is 15-15, the energy consumption is minimum. Because

2000 Energy Saving 1800 Screen Off Prevew with 1600 Screen On Minimum Phone Lifting Arm Frame Rate Rotating 1400 Turn off Camera Mationless Adjust Screen Walking Brizhtness based Up Preview with Fine 1200 Turn Offnior on Ambient Light uning 1000 without Low the 176*144320*240720°4809607201280720 Shooting for a Laving Arn Brightness to 5 Down Shoot Long Time ame Rate Fig.13. Energy Consumption with Different Resolution of Sony LT26w Maximum Medium Minimum Fig.15. Energy Saving Schemes Corresponding to Different States IV.SYSTEM EVALUATION We implement our system on Samsung GT-19250 smart- phone running on Google's Android platform.The version of the Android system is 4.4.2.We use Monsoon power monitor to measure the power consumption of the phone. (a)Different Frame Rate of Preview (b)Different Brightness of Screen Fig.14.Energy Consumption of Different Brightness and Preview Frame Rates A.Impact of Sensors'Sampling Rate 1)Energy Consumption of Sensors with Various Sampling Rates:The sensors'maximum sampling rate of Samsung GT- of Android mechanism,the phone adaptively chooses the 19250 is 100 Hz.We vary the sampling rate in twenty levels suitable preview frame rate in the range by itself.Therefore with step of 5.Energy consumption of sensors are different the energy consumption of last two situations are same. with various sampling rates as shown in Figure 16(a).We Observation of Screen.Brightness of screen is related to observe that power consumed is relatively big after 25 Hz. the energy consumption shown in Figures 14(b).The range of 2)Energy Consumption of Calculation of Sensors'Data: brightness is 0-255.0 stands for darkest and 255 for brightest. Energy consumption of calculation is related to the data size Once the brightness drops,the energy consumption decreases. With the sampling rate increasing,the energy also increases. 2)Energy Saving Scheme:For obtained states,we apply We observe that the power of calculation of sliding window is corresponding energy saving strategies,as shown in Figure 15 only 1.5 mW.SVM model consumes about 4.34 mW as it is only used to do prediction with a trained model.The power If obtained state is in body level,the screen and the camera of FFT is about 122 mW.The energy of other calculation can will be turned off as the user doesn't need to look at the screen. be ignored.Therefore,only energy of FFT is needed to be Further more,if the states always belong to body level for a considered.But compared to the power of camera which is long time(15 minutes is chosen in our implement),the sensors about 1500 mW,it is unobservable. will be turned off until the camera software is used again. 3)Activity Recognition Accuracies with Various Sampling If obtained state is in arm level,the screen is turned on Rates:We use six sampling rates,which are 2 samples/second. and its brightness is adjusted based on ambient light as lifting 5 samples/second,10 samples/second,20 samples/second,50 up your arm.The brightness will be declined to 5 as laying samples/second and 100 samples/second,to evaluate the im- down your arm.The brightness is set to five levels according pact of sampling rate of sensors on the performance of recog- to different environment shown in Table II. nition accuracy.In Figures 16(b),the accuracies of actions in body level are showed.For motionlessness,the accuracy has TABLE II. THE BRIGHTNESS OF SCREEN IN DIFFERENT ENVIRONMENT no relationship with the sampling rate.For walk,the accuracy is low with rate of 2 Hz and bigger than 90%with the other Number Environment Ambient Light (SI lux) of Screen five rates.For jog,the accuracy is bigger than 90%at the rate Day.outdoor,sunny >6000 180 of 10 Hz.In Figures 16(c),the accuracies of actions in arm Day,outdoor,cloudy 5006000T 130 level are showed.We can find out that the accuracy of 20 Hz Day,indoor,no lamp 100300 80 is 100%.In Figures 16(d).the accuracies of actions in wrist Night,outdoor,street lamp 100 55 Night,indoor,lamp 300500 105 level are showed and it is 100%when the sample rate is 100 Hz. If obtained state is in wrist level,the camera will be turned on and stay in preview mode.When you rotate the phone, B.Trade off between Energy Consumption and Recognition Accuracy camera is set to work with smallest frame rate supported by the phobe.In fine-tuning state,camera works with increased From Figures 16(b)(c)(d).the accuracy can be 100%with frame rate (median value is used).And in shooting state,all 100 Hz.But it not energy efficiency as shown in Figure the indexes will return to normal.All the parameters can be 16(a).Therefore,we need make a trade off between energy changed by the user if the parameters do not fit them consumption and recognition accuracy

176*144 320*240 720*480 960*720 1280*720 Power (mW) 1000 1200 1400 1600 1800 2000 Fig. 13. Energy Consumption with Different Resolution of Sony LT26w 15-15 15-30 24-30 Power of Preview (mW) 660 680 700 720 740 760 780 800 (a) Different Frame Rate of Preview Brightness of Screen 255 230 205 180 155 130 105 80 55 30 5 Power of Screen(mW) 400 500 600 700 800 900 1000 1100 1200 1300 1400 (b) Different Brightness of Screen Fig. 14. Energy Consumption of Different Brightness and Preview Frame Rates of Android mechanism, the phone adaptively chooses the suitable preview frame rate in the range by itself. Therefore the energy consumption of last two situations are same. Observation of Screen. Brightness of screen is related to the energy consumption shown in Figures 14(b). The range of brightness is 0-255. 0 stands for darkest and 255 for brightest. Once the brightness drops, the energy consumption decreases. 2) Energy Saving Scheme: For obtained states, we apply corresponding energy saving strategies, as shown in Figure 15. If obtained state is in body level, the screen and the camera will be turned off as the user doesn’t need to look at the screen. Further more, if the states always belong to body level for a long time (15 minutes is chosen in our implement), the sensors will be turned off until the camera software is used again. If obtained state is in arm level, the screen is turned on and its brightness is adjusted based on ambient light as lifting up your arm. The brightness will be declined to 5 as laying down your arm. The brightness is set to five levels according to different environment shown in Table II. TABLE II. THE BRIGHTNESS OF SCREEN IN DIFFERENT ENVIRONMENT Number Environment Ambient Brightness Light (SI lux) of Screen 1 Day, outdoor, sunny >6000 180 2 Day, outdoor, cloudy 500∼6000 130 3 Day, indoor, no lamp 100∼300 80 4 Night, outdoor, street lamp <100 55 5 Night, indoor, lamp 300∼500 105 If obtained state is in wrist level, the camera will be turned on and stay in preview mode. When you rotate the phone, camera is set to work with smallest frame rate supported by the phobe. In fine-tuning state, camera works with increased frame rate (median value is used). And in shooting state, all the indexes will return to normal. All the parameters can be changed by the user if the parameters do not fit them. Maximum Medium Minimum Screen Off Turn Off Sensors without Shooting for a Long Time Screen On Adjust Screen Brightness based on Ambient Light Preview with Minimum Frame Rate Turn off Camera Energy Saving Motionless Walking, Jogging, Low the Brightness to 5 Lifting Arm Up Laying Arm Down Phone Rotating Preview with Median Frame Rate Fine￾tuning Preview with Normal Frame Rate Shooting Fig. 15. Energy Saving Schemes Corresponding to Different States IV. SYSTEM EVALUATION We implement our system on Samsung GT-I9250 smart￾phone running on Google’s Android platform. The version of the Android system is 4.4.2. We use Monsoon power monitor to measure the power consumption of the phone. A. Impact of Sensors’ Sampling Rate 1) Energy Consumption of Sensors with Various Sampling Rates: The sensors’ maximum sampling rate of Samsung GT￾I9250 is 100 Hz. We vary the sampling rate in twenty levels with step of 5. Energy consumption of sensors are different with various sampling rates as shown in Figure 16(a). We observe that power consumed is relatively big after 25 Hz. 2) Energy Consumption of Calculation of Sensors’ Data: Energy consumption of calculation is related to the data size. With the sampling rate increasing, the energy also increases. We observe that the power of calculation of sliding window is only 1.5 mW. SVM model consumes about 4.34 mW as it is only used to do prediction with a trained model. The power of FFT is about 122 mW. The energy of other calculation can be ignored. Therefore, only energy of FFT is needed to be considered. But compared to the power of camera which is about 1500 mW, it is unobservable. 3) Activity Recognition Accuracies with Various Sampling Rates: We use six sampling rates, which are 2 samples/second, 5 samples/second, 10 samples/second, 20 samples/second, 50 samples/second and 100 samples/second, to evaluate the im￾pact of sampling rate of sensors on the performance of recog￾nition accuracy. In Figures 16(b), the accuracies of actions in body level are showed. For motionlessness, the accuracy has no relationship with the sampling rate. For walk, the accuracy is low with rate of 2 Hz and bigger than 90% with the other five rates. For jog, the accuracy is bigger than 90% at the rate of 10 Hz. In Figures 16(c), the accuracies of actions in arm level are showed. We can find out that the accuracy of 20 Hz is 100%. In Figures 16(d), the accuracies of actions in wrist level are showed and it is 100% when the sample rate is 100 Hz. B. Trade off between Energy Consumption and Recognition Accuracy From Figures 16(b)(c)(d), the accuracy can be 100% with 100 Hz. But it not energy efficiency as shown in Figure 16(a). Therefore, we need make a trade off between energy consumption and recognition accuracy

180 160 0.9 140 02 01 er Gravity Senso Lay Dow (a) (b) (c) 0.95 0.8 07 0.9 g0.5 0.3 0.65 g tpy的 (d) (e) (① Fig.16.Evaluation of Accuracy.(a)Sensors'sampling rates and corresponding energy consumption.(b)Sensors'sampling rates and accuracies of body level. (c)Sensors'sampling rates and accuracies of arm level.(d)Sensors'sampling rates and accuracies of wrist level.(e)Confusion matrix for eight actions.(f) Accuracies of different users. In Figures 16(b),energy doesn't increase a lot when sampling rate changes from 10 Hz to 20 Hz while the accuracy of three actions can be increased to 94%.And it is similar to the actions in arm movement as the accuracy can be 100% at 20 Hz.In Figures 16(d),the accuracy is good when the sampling rate is 20 Hz.And the accuracy doesn't increase a lot if the sampling rate changes from 20 Hz to 50 Hz which will result in one times more energy consumption.Therefore 20 Hz is used as sampling rate to recognize all the actions. (a)10 Shoots in 5 Minutes in Out-(b)10 Shoots in 5 Minutes in 5 door Using 3 Schemes Different Environment C.Recognition Accuracy Fig.17.Energy Consumption of the Process of Shooting using Different 1)Recognition Accuracy of Our Scheme:The average Schemes and in Different Environment accuracy of our scheme is 95.5%.Figure 16(e)plots the confusion matrix for eight actions with the sampling rate of 20 Hz.Each row denotes the actual actions performed by the randomly in the same environment outdoor (cloudy)and the user and each column the actions it was classified into.Each result is showed in Figures 17(a).Taking advantage of our element in the matrix corresponds to the fraction of action in scheme,energy consumption can be reduced 46.5%compared the row that were classified as the action in the column to no scheme and 12.7%compared to turn on/off scheme. 2)Recognition Accuracy of Different People:In order to Then,we make measures when we use the smart camera evaluate the feasibility,we invite 5 users to test our smart phone in five different environment,as shown in Table II camera in different environment.All the users use the phone In 17(b),the x-label respectively maps to the environment. to take 10 photos in five minutes.During the process,the Using our scheme,the energy consumption will change with users may perform any actions of three levels.The average the transformation of environment.Compared to turn on/off accuracies of all ten processes are showed in Figures 16(f). scheme,energy consumption decreases by 8%,12.7%,14%, We can find out that all the accuracies are above 85%and two 14.9%and 13.3%respectively. of them are above 90%. V.RELATED WORK D.Energy Consumption A.Energy Saving We measure the energy consumption under three schemes, Prior work on energy saving of smart-phone can be classi- which are no scheme,turn on/off scheme and our context- fied into three parts,analysis of hardwares'energy consump- aware energy-saving scheme.We take 10 shoots in 5 minutes tion [4],[5],[6],[7],power model [8][9]and energy saving

Linear Accelerometer Gravity Sensor Gyroscope Power of sensors (mW) 0 20 40 60 80 100 120 140 160 180 5 Hz 10 Hz 15 Hz 20 Hz 25 Hz 30 Hz 35 Hz 40 Hz 45 Hz 50 Hz 55 Hz 60 Hz 65 Hz 70 Hz 75 Hz 80 Hz 85 Hz 90 Hz 95 Hz 100 Hz (a) Motionless Walking Jogging Accuracy 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 Hz 5 Hz 10 Hz 20 Hz 50 Hz 100 Hz (b) Lift Up Lay Down Accuracy 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 Hz 5 Hz 10 Hz 20 Hz 50 Hz 100 Hz (c) Phone Rotate Fine-tuning Shooting Accuracy 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 Hz 5 Hz 10 Hz 20 Hz 50 Hz 100 Hz (d) 1 0 0 0 0 0 0 0 0 1 0.06 0 0 0 0 0 0 0 0.94 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0.89 0 0 0 0 0 0 0 0.11 0.90 0.09 0 0 0 0 0 0 0.10 0.91 motionlesswalk jog lift uplay downrotatefine-tuneshoot motionless walk jog lift up lay down rotate fine-tune shoot (e) U1 U2 U3 U4 U5 Accuracy 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 (f) Fig. 16. Evaluation of Accuracy. (a) Sensors’ sampling rates and corresponding energy consumption. (b) Sensors’ sampling rates and accuracies of body level. (c) Sensors’ sampling rates and accuracies of arm level. (d) Sensors’ sampling rates and accuracies of wrist level. (e) Confusion matrix for eight actions. (f) Accuracies of different users. In Figures 16(b), energy doesn’t increase a lot when sampling rate changes from 10 Hz to 20 Hz while the accuracy of three actions can be increased to 94%. And it is similar to the actions in arm movement as the accuracy can be 100% at 20 Hz. In Figures 16(d), the accuracy is good when the sampling rate is 20 Hz. And the accuracy doesn’t increase a lot if the sampling rate changes from 20 Hz to 50 Hz which will result in one times more energy consumption. Therefore, 20 Hz is used as sampling rate to recognize all the actions. C. Recognition Accuracy 1) Recognition Accuracy of Our Scheme: The average accuracy of our scheme is 95.5%. Figure 16(e) plots the confusion matrix for eight actions with the sampling rate of 20 Hz. Each row denotes the actual actions performed by the user and each column the actions it was classified into. Each element in the matrix corresponds to the fraction of action in the row that were classified as the action in the column. 2) Recognition Accuracy of Different People: In order to evaluate the feasibility, we invite 5 users to test our smart camera in different environment. All the users use the phone to take 10 photos in five minutes. During the process, the users may perform any actions of three levels. The average accuracies of all ten processes are showed in Figures 16(f). We can find out that all the accuracies are above 85% and two of them are above 90%. D. Energy Consumption We measure the energy consumption under three schemes, which are no scheme, turn on/off scheme and our context￾aware energy-saving scheme. We take 10 shoots in 5 minutes No Scheme Turn on/off Scheme Our Scheme Energy Consumption (uAh) #104 2.5 3 3.5 4 4.5 5 5.5 6 (a) 10 Shoots in 5 Minutes in Out￾door Using 3 Schemes 1 2 3 4 5 Energy Consumption (uAh) #104 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 turn on/off scheme our scheme (b) 10 Shoots in 5 Minutes in 5 Different Environment Fig. 17. Energy Consumption of the Process of Shooting using Different Schemes and in Different Environment randomly in the same environment outdoor (cloudy) and the result is showed in Figures 17(a). Taking advantage of our scheme, energy consumption can be reduced 46.5% compared to no scheme and 12.7% compared to turn on/off scheme. Then, we make measures when we use the smart camera phone in five different environment, as shown in Table II. In 17(b), the x-label respectively maps to the environment. Using our scheme, the energy consumption will change with the transformation of environment. Compared to turn on/off scheme, energy consumption decreases by 8%, 12.7%, 14%, 14.9% and 13.3% respectively. V. RELATED WORK A. Energy Saving Prior work on energy saving of smart-phone can be classi- fied into three parts, analysis of hardwares’ energy consump￾tion [4], [5], [6], [7], power model [8][9] and energy saving

schemes.Chen et al.10 analyze the power consumption ACKNOWLEDGMENT of AMOLED displays in multimedia applications.Camera recoding incurs high power cost.LiKamWa R et al.[11]do re- This work is supported in part by National Natural Science search on the image sensor and reveal two energy-proportional Foundation of China under Grant Nos.61472185,61373129 mechanisms which are supported by current image sensor but 61321491,91218302;Key Project of Jiangsu Research Pro- unused in mobile system.It indeed saves energy.But it only gram under Grant No.BE2013116:EU FP7 IRSES Mobile- focuses on the energy consumption of the moment of shooting, Cloud Project under Grant No.612212.This work is par- while overlooking the consumption of preparation.Han et al. tially supported by Collaborative Innovation Center of Novel [12]study the energy cost made by human-screen interaction Software Technology and Industrialization.Lei Xie is the such as scrolling on screen.They propose a scrolling-speed- corresponding author. adaptive frame rate controlling system to save energy.Dietrich et al.[3]detect the game's current state and lower the proces- REFERENCES sor's voltage and frequency whenever possible to save energy. [1]"KS Mobile,"http://www.cmcm.com/zh-cn/. [2] "Monsoon PowerMonitor,"https://www.msoon.com/LabEquipment/ B.Activity Sensing PowerMonitor/. With the development of phone's built-in sensors,various [31 B.Dietrich and S.Chakraborty,"Power management using game state detection on android smartphones,"in Proc.of ACM MobiSys,2013. approaches of activity recognition have been done.They can be classified into single-sensor and multi-sensors sensing. [4]X.Fan,W.-D.Weber,and L.A.Barroso,"Power provisioning for a warehouse-sized computer,"in Proc.of ACM S/GARCH,2007 Single sensor is used in the following work.Built-in [5] F.Bellosa.A.Weissel.M.Waitz.and S.Kellner."Event-driven energy microphone is used to detect the events that are closely related accounting for dynamic thermal management,"in Proc.of CoLp,2003. to sleep quality,including body movement,couch and snore [6 D.Rajan,R.Zuck,and C.Poellabauer,"Workload-aware dual-speed [13].Using built-in accelerometers,user's daily activities such dynamic voltage scaling,"in Proc.of /EEE RTCSA.2006. as walking,jogging,upstairs,downstairs,sitting and standing [7] N.Balasubramanian,A.Balasubramanian,and A.Venkataramani, are recognized in [14].With the labeled accelerometer data, "Energy consumption in mobile phones:a measurement study and implications for network applications,"in Proc.of ACM S/GCOMM, they apply four machine learning algorithms and make some 2009 analysis.Lee et al.[15]use accelerometers with hierarchical [8]M.Dong and L.Zhong,"Self-constructive high-rate system energy hidden markov models to distinguish the daily actions. modeling for battery-powered mobile systems,"in Proc.of ACM MobiSys,2011. Multi-sensors are used in the following work.Shahzad et [9] F.Xu.Y.Liu,Q.Li.and Y.Zhang."V-edge:Fast self-constructive al.[16]propose a gesture based user authentication scheme power modeling of smartphones based on battery voltage dynamics."in for the secure unlocking of touch screen devices.It makes Proc.of NSDI,2013. use of the coordinates of each touch point on the screen, 10 X.Chen,Y.Chen,Z.Ma,and F.C.Fernandes,"How is energy accelerometer values and time stamps.Chen et al.[17]take consumed in smartphone display applications?in Proc.of ACM advantage of features as light,phone situation,stationary and HotMobile.2013. silence to monitor user's sleep.They need to use several [11] R.LiKamWa,B.Priyantha,M.Philipose,L.Zhong,and P.Bahl different sensors to obtain all the phone's information.Driving "Energy characterization and optimization of image sensing toward style,which is concerned with man's life,is recognized by continuous mobile vision,"in Proc.of ACM MobiSys,2013. [12] H.Han,J.Yu.H.Zhu,Y.Chen,J.Yang.G.Xue,Y.Zhu,and M.Li,"E using the gyroscope,accelerometer,magnetometer,GPS and 3:energy-efficient engine for frame rate adaptation on smartphones,"in video[18].Bo et al.[19]propose a framework to verify whether Proc.of ACM Sensys,2013. the current user is the legitimate owner of the smartphone [13] T.Hao,G.Xing,and G.Zhou,"isleep:unobtrusive sleep quality based on the behavioral biometrics,including touch behaviors monitoring using smartphones,"in Proc.of ACM Sensys,2013. and walking pattens.These features are extracted from smart- [14] J.R.Kwapisz,G.M.Weiss,and S.A.Moore,"Activity recognition phone's built-in accelerometer and gyroscope. using cell phone accelerometers,"SIGKDD,vol.12,no.2,pp.74-82. 2011. VI.CONCLUSION AND FUTURE WORK [15) Y.-S.Lee and S.-B.Cho."Activity recognition using hierarchical hidden markov models on a smartphone with 3d accelerometer."in Proc.of In this paper,we propose a context aware energy-saving Springer HAIS,2011. scheme for smart camera phone based on activity sensing. [16 M.Shahzad,A.X.Liu,and A.Samuel,"Secure unlocking of mobile We take advantage of the features of activities and maintain touch screen devices by simple gestures:You can see it but you can an activity state machine to do the recognition.Then energy not do it,"in Proc.of ACM MobiCom.2013. saving scheme is applied based on the result of recognition [17]Z.Chen,M.Lin,F.Chen,N.D.Lane,G.Cardone,R.Wang,T.Li Our solution can perceive the user's activities with an average Y.Chen,T.Choudhury,and A.T.Campbell,"Unobtrusive sleep monitoring using smartphones,"in Proc.of IEEE PervasiveHealth, accuracy of 95.5%and reduce the overall energy consumption 2013. by 46.5%for smart camera phones. [18]D.A.Johnson and M.M.Trivedi,"Driving style recognition using a Following the current research,there are three possible smartphone as a sensor platform,"in Proc.of IEEE ITSC,2011. directions for future work.First,more data of the process [19]C.Bo,L.Zhang.X.-Y.Li,Q.Huang,and Y.Wang."Silentsense:silent user identification via touch and movement behavioral biometrics."in can be collected with our work to improve the design and Proc.of ACM MobiCom,2013. implementation.Second,a self-constructive user preference learning can be designed to automatically extract the user perference of software settings.Third,to the phone whose configuration is too low,more simple strategy can be designed to avoid the possible delay for changing modes

schemes. Chen et al. [10] analyze the power consumption of AMOLED displays in multimedia applications. Camera recoding incurs high power cost. LiKamWa R et al. [11] do re￾search on the image sensor and reveal two energy-proportional mechanisms which are supported by current image sensor but unused in mobile system. It indeed saves energy. But it only focuses on the energy consumption of the moment of shooting, while overlooking the consumption of preparation. Han et al. [12] study the energy cost made by human-screen interaction such as scrolling on screen. They propose a scrolling-speed￾adaptive frame rate controlling system to save energy. Dietrich et al. [3] detect the game’s current state and lower the proces￾sor’s voltage and frequency whenever possible to save energy. B. Activity Sensing With the development of phone’s built-in sensors, various approaches of activity recognition have been done. They can be classified into single-sensor and multi-sensors sensing. Single sensor is used in the following work. Built-in microphone is used to detect the events that are closely related to sleep quality, including body movement, couch and snore [13]. Using built-in accelerometers, user’s daily activities such as walking, jogging, upstairs, downstairs, sitting and standing are recognized in [14]. With the labeled accelerometer data, they apply four machine learning algorithms and make some analysis. Lee et al. [15] use accelerometers with hierarchical hidden markov models to distinguish the daily actions. Multi-sensors are used in the following work. Shahzad et al. [16] propose a gesture based user authentication scheme for the secure unlocking of touch screen devices. It makes use of the coordinates of each touch point on the screen, accelerometer values and time stamps. Chen et al. [17] take advantage of features as light, phone situation, stationary and silence to monitor user’s sleep. They need to use several different sensors to obtain all the phone’s information. Driving style, which is concerned with man’s life, is recognized by using the gyroscope, accelerometer, magnetometer, GPS and video[18]. Bo et al. [19] propose a framework to verify whether the current user is the legitimate owner of the smartphone based on the behavioral biometrics, including touch behaviors and walking pattens. These features are extracted from smart￾phone’s built-in accelerometer and gyroscope. VI. CONCLUSION AND FUTURE WORK In this paper, we propose a context aware energy-saving scheme for smart camera phone based on activity sensing. We take advantage of the features of activities and maintain an activity state machine to do the recognition. Then energy saving scheme is applied based on the result of recognition. Our solution can perceive the user’s activities with an average accuracy of 95.5% and reduce the overall energy consumption by 46.5% for smart camera phones. Following the current research, there are three possible directions for future work. First, more data of the process can be collected with our work to improve the design and implementation. Second, a self-constructive user preference learning can be designed to automatically extract the user perference of software settings. Third,to the phone whose configuration is too low, more simple strategy can be designed to avoid the possible delay for changing modes. ACKNOWLEDGMENT This work is supported in part by National Natural Science Foundation of China under Grant Nos. 61472185,61373129, 61321491, 91218302; Key Project of Jiangsu Research Pro￾gram under Grant No. BE2013116; EU FP7 IRSES Mobile￾Cloud Project under Grant No. 612212. This work is par￾tially supported by Collaborative Innovation Center of Novel Software Technology and Industrialization. Lei Xie is the corresponding author. REFERENCES [1] “KS Mobile,” http://www.cmcm.com/zh-cn/. [2] “Monsoon PowerMonitor,” https://www.msoon.com/LabEquipment/ PowerMonitor/. [3] B. Dietrich and S. Chakraborty, “Power management using game state detection on android smartphones,” in Proc. of ACM MobiSys, 2013. [4] X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a warehouse-sized computer,” in Proc. of ACM SIGARCH, 2007. [5] F. Bellosa, A. Weissel, M. Waitz, and S. Kellner, “Event-driven energy accounting for dynamic thermal management,” in Proc. of COLP, 2003. [6] D. Rajan, R. Zuck, and C. Poellabauer, “Workload-aware dual-speed dynamic voltage scaling,” in Proc. of IEEE RTCSA, 2006. [7] N. Balasubramanian, A. Balasubramanian, and A. Venkataramani, “Energy consumption in mobile phones: a measurement study and implications for network applications,” in Proc. of ACM SIGCOMM, 2009. [8] M. Dong and L. Zhong, “Self-constructive high-rate system energy modeling for battery-powered mobile systems,” in Proc. of ACM MobiSys, 2011. [9] F. Xu, Y. Liu, Q. Li, and Y. Zhang, “V-edge: Fast self-constructive power modeling of smartphones based on battery voltage dynamics.” in Proc. of NSDI, 2013. [10] X. Chen, Y. Chen, Z. Ma, and F. C. Fernandes, “How is energy consumed in smartphone display applications?” in Proc. of ACM HotMobile, 2013. [11] R. LiKamWa, B. Priyantha, M. Philipose, L. Zhong, and P. Bahl, “Energy characterization and optimization of image sensing toward continuous mobile vision,” in Proc. of ACM MobiSys, 2013. [12] H. Han, J. Yu, H. Zhu, Y. Chen, J. Yang, G. Xue, Y. Zhu, and M. Li, “E 3: energy-efficient engine for frame rate adaptation on smartphones,” in Proc. of ACM Sensys, 2013. [13] T. Hao, G. Xing, and G. Zhou, “isleep: unobtrusive sleep quality monitoring using smartphones,” in Proc. of ACM Sensys, 2013. [14] J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Activity recognition using cell phone accelerometers,” SIGKDD, vol. 12, no. 2, pp. 74–82, 2011. [15] Y.-S. Lee and S.-B. Cho, “Activity recognition using hierarchical hidden markov models on a smartphone with 3d accelerometer,” in Proc. of Springer HAIS, 2011. [16] M. Shahzad, A. X. Liu, and A. Samuel, “Secure unlocking of mobile touch screen devices by simple gestures: You can see it but you can not do it,” in Proc.of ACM MobiCom, 2013. [17] Z. Chen, M. Lin, F. Chen, N. D. Lane, G. Cardone, R. Wang, T. Li, Y. Chen, T. Choudhury, and A. T. Campbell, “Unobtrusive sleep monitoring using smartphones,” in Proc. of IEEE PervasiveHealth, 2013. [18] D. A. Johnson and M. M. Trivedi, “Driving style recognition using a smartphone as a sensor platform,” in Proc. of IEEE ITSC, 2011. [19] C. Bo, L. Zhang, X.-Y. Li, Q. Huang, and Y. Wang, “Silentsense: silent user identification via touch and movement behavioral biometrics,” in Proc. of ACM MobiCom, 2013