cific human activity.CARM uses this quantitative correlation as the The third challenge is that CSI values are too noisy to be dir- profiling mechanism and recognizes a given activity by matching it ectly used for human activity recognition.Even in a static en- to the best-fit profile. vironment without any human activity,CSI values fluctuate be- cause WiFi devices are susceptible to surrounding electromagnetic noises.Moreover,the internal state changes in WiFi devices,e.g.. transmission rate adaptation and transmission power adaptation of- Wireless route ten introduce impulse and burst noises in CSI values.General pur- pose denoising methods,such as low-pass filters or median filters, do not perform well in removing these impulse and bursty noises for two reasons:First,the sampling rates that these methods require are much higher than the frequency of the WiFi signal.Second,the noise density in CSI values is too high for traditional filters,which Figure 1:CARM System only work well for low density noise.In this paper,we propose a Our CSI-speed model and CSI-activity model advance the state- principal component analysis(PCA)based CSI denoising scheme. of-the-art on WiFi signal based human activity recognition from This scheme is based on our observation that the signal fluctuations two fronts.First,they provide us the theoretical basis to under- caused by body movements in all subcarriers of the CSI values are stand,even quantitatively.the relationship between CSI value dy correlated. namics and human movement speeds,and the relationship between The fourth challenge is to capture body movements in the pres- the movement speeds of different human body parts and human ence of carrier frequency offset(CFO).CFO is the dynamically activities.Regarding the relationship between CSI value dynamics changing difference in carrier frequencies between a pair of WiFi and human movement speeds,for example,our model shows that devices,which occurs due to the minor physical differences in high-speed body part movement generates high-frequency changes hardware and other factors such as temperature changes [8].CFO in CSI values.Regarding the relationship between the movement causes the phase values of the received signal to change,making it speeds of different human body parts and human activities,tak- hard to distinguish whether the phase value changed is due to CFO ing the activity of falling down as an example,our model shows or due to human movement.To address this challenge.we use the that it can be characterized as a sudden increase in body movement CSI signal power to infer the body movement.We show that CSI speed in less than one second.Second,these two models provide us signal power is not affected by CFO,but retains information about the tunable parameters to optimize the performance of WiFi signal the movement speeds of the body. based human activity recognition.For example,according to our The fifth challenge is to automatically detect the start and end models,the CSI sampling rate should be chosen as 800 samples per of a human activity.To address this challenge,we use the eigen- second because the typical human movement speed corresponds to vectors obtained from PCA.The key idea is that in the absence of CSI components of lower than 300 Hz. any activity,the time-series of CSI values contain random noise and consequently,the signal eigenvector varies randomly.During a hu- 1.4 Technical Challenges and Our Solutions man activity,the signals in subcarriers become correlated and the The first technical challenge is to estimate human movement signal eigenvector becomes smooth.We capture the smoothness of speeds from CSI values based on our CSI-speed model.This the eigenvector by calculating its high-frequency energy and com is challenging because the CSI measurements at the receiver are pare it to a dynamically adapting threshold to detect start and end. mixed WiFi signals arrived from multiple paths,which changes as 1.5 Key Technical Novelty and Results human moves.Furthermore,different human body parts move at different speeds for a given activity and the WiFi signals reflected The key technical novelty of this paper is two fold.First,we pro- by different body parts are also mixed at the receiver.Our key ob- pose the CSI-speed model and the CSI-activity model to quantify servation is that these signals are linearly combined so that their the correlation between CSI value dynamics and a specific human frequencies are preserved when they are mixed together.There- activity.Second,we propose a set of signal processing techniques. fore,we use Discrete Wavelet Transform (DWT)to separate the such as PCA based denoising and DWT based feature extraction, for human activity recognition based on the CSI-speed model and frequency components that represent different movement speeds. The advantage of DWT is that it provides a proper tradeoff between the CSI-activity model.The key technical depth of this paper lies time and frequency resolution and enables the measurement of both in the signal processing aspect such as the theoretical analysis of the correlation between CSI values of subcarriers and the relation- fast and slow activities. The second challenge is to build the CSI-activity model that is ro- ship between multi-path speeds and CFR power.We implemented bust for different humans.For the same activity,to a certain degree, CARM on commercial WiFi devices and evaluated it in multiple environments.Our results show that CARM achieves an average different people perform it differently and even the same person performs it differently at different times.To address this challenge, activity recognition accuracy of 96%.For a new environment and we propose a Hidden Markov Model(HMM)based human activity a new person that the system has never been trained on,CARM can recognition approach.We use the patterns of movement speeds for still achieve a recognition accuracy for more than 80% different activities to build their corresponding HMM based mod- els.The features that we extract to infer the speed patterns are 2.RELATED WORK only affected by movement speeds of the body and are relatively Existing work on device-free human activity recognition and agonistic to environmental changes.This enables us to recognize localization can be divided into four categories:Received Signal activities even when the environment changes.We choose HMM Strength Indicator(RSSI)based,specialized hardware based.radar because of its inherent capability to recognize the same activities based,and CSI based. that are done at different speeds.To recognize a sample of an un- RSSI Based:RSSI based human activity recognition systems known activity,we evaluate the unknown samples against HMMs leverage the signal strength changes caused by human activities of all activities and find the model that gives the highest likelihood. [3,22,23].This approach can only do coarse grained human activitycific human activity. CARM uses this quantitative correlation as the profiling mechanism and recognizes a given activity by matching it to the best-fit profile. Wireless router Laptop Wireless signal reflection Figure 1: CARM System Our CSI-speed model and CSI-activity model advance the state￾of-the-art on WiFi signal based human activity recognition from two fronts. First, they provide us the theoretical basis to under￾stand, even quantitatively, the relationship between CSI value dy￾namics and human movement speeds, and the relationship between the movement speeds of different human body parts and human activities. Regarding the relationship between CSI value dynamics and human movement speeds, for example, our model shows that high-speed body part movement generates high-frequency changes in CSI values. Regarding the relationship between the movement speeds of different human body parts and human activities, tak￾ing the activity of falling down as an example, our model shows that it can be characterized as a sudden increase in body movement speed in less than one second. Second, these two models provide us the tunable parameters to optimize the performance of WiFi signal based human activity recognition. For example, according to our models, the CSI sampling rate should be chosen as 800 samples per second because the typical human movement speed corresponds to CSI components of lower than 300 Hz. 1.4 Technical Challenges and Our Solutions The first technical challenge is to estimate human movement speeds from CSI values based on our CSI-speed model. This is challenging because the CSI measurements at the receiver are mixed WiFi signals arrived from multiple paths, which changes as human moves. Furthermore, different human body parts move at different speeds for a given activity and the WiFi signals reflected by different body parts are also mixed at the receiver. Our key ob￾servation is that these signals are linearly combined so that their frequencies are preserved when they are mixed together. There￾fore, we use Discrete Wavelet Transform (DWT) to separate the frequency components that represent different movement speeds. The advantage of DWT is that it provides a proper tradeoff between time and frequency resolution and enables the measurement of both fast and slow activities. The second challenge is to build the CSI-activity model that is ro￾bust for different humans. For the same activity, to a certain degree, different people perform it differently and even the same person performs it differently at different times. To address this challenge, we propose a Hidden Markov Model (HMM) based human activity recognition approach. We use the patterns of movement speeds for different activities to build their corresponding HMM based mod￾els. The features that we extract to infer the speed patterns are only affected by movement speeds of the body and are relatively agonistic to environmental changes. This enables us to recognize activities even when the environment changes. We choose HMM because of its inherent capability to recognize the same activities that are done at different speeds. To recognize a sample of an un￾known activity, we evaluate the unknown samples against HMMs of all activities and find the model that gives the highest likelihood. The third challenge is that CSI values are too noisy to be dir￾ectly used for human activity recognition. Even in a static en￾vironment without any human activity, CSI values fluctuate be￾cause WiFi devices are susceptible to surrounding electromagnetic noises. Moreover, the internal state changes in WiFi devices, e.g., transmission rate adaptation and transmission power adaptation of￾ten introduce impulse and burst noises in CSI values. General pur￾pose denoising methods, such as low-pass filters or median filters, do not perform well in removing these impulse and bursty noises for two reasons: First, the sampling rates that these methods require are much higher than the frequency of the WiFi signal. Second, the noise density in CSI values is too high for traditional filters, which only work well for low density noise. In this paper, we propose a principal component analysis (PCA) based CSI denoising scheme. This scheme is based on our observation that the signal fluctuations caused by body movements in all subcarriers of the CSI values are correlated. The fourth challenge is to capture body movements in the pres￾ence of carrier frequency offset (CFO). CFO is the dynamically changing difference in carrier frequencies between a pair of WiFi devices, which occurs due to the minor physical differences in hardware and other factors such as temperature changes [8]. CFO causes the phase values of the received signal to change, making it hard to distinguish whether the phase value changed is due to CFO or due to human movement. To address this challenge, we use the CSI signal power to infer the body movement. We show that CSI signal power is not affected by CFO, but retains information about the movement speeds of the body. The fifth challenge is to automatically detect the start and end of a human activity. To address this challenge, we use the eigen￾vectors obtained from PCA. The key idea is that in the absence of any activity, the time-series of CSI values contain random noise and consequently, the signal eigenvector varies randomly. During a hu￾man activity, the signals in subcarriers become correlated and the signal eigenvector becomes smooth. We capture the smoothness of the eigenvector by calculating its high-frequency energy and com￾pare it to a dynamically adapting threshold to detect start and end. 1.5 Key Technical Novelty and Results The key technical novelty of this paper is two fold. First, we pro￾pose the CSI-speed model and the CSI-activity model to quantify the correlation between CSI value dynamics and a specific human activity. Second, we propose a set of signal processing techniques, such as PCA based denoising and DWT based feature extraction, for human activity recognition based on the CSI-speed model and the CSI-activity model. The key technical depth of this paper lies in the signal processing aspect such as the theoretical analysis of the correlation between CSI values of subcarriers and the relation￾ship between multi-path speeds and CFR power. We implemented CARM on commercial WiFi devices and evaluated it in multiple environments. Our results show that CARM achieves an average activity recognition accuracy of 96%. For a new environment and a new person that the system has never been trained on, CARM can still achieve a recognition accuracy for more than 80%. 2. RELATED WORK Existing work on device-free human activity recognition and localization can be divided into four categories: Received Signal Strength Indicator (RSSI) based, specialized hardware based, radar based, and CSI based. RSSI Based: RSSI based human activity recognition systems leverage the signal strength changes caused by human activities [3,22,23]. This approach can only do coarse grained human activity