Table 1:Label Description eter and gyroscope data among 5 different users.Then we use them to train the classify model.We perform the clas- UST going upstairs sifier on the previous 5 users and 3 new users.They go DST going downstairs up/downstairs,walk or take elevator.Figure 3(a)shows the WALK walking accuracy of Activity Classifier on the collected data.Particu- EHG hypergravity in elevator. larly,we don't evaluate the EUP/EDOWN because that EUP EWL weightlessness in elevator. is the combination of EHG and EWL,and EDOWN is the E UP elevator going up combination of EWL and EHG.On average,we achieve a E_DOWN elevator going down classification accuracy of 96.2%. SS stand still Location Accuracy.To evaluate the location performance of FootStep-Tracker,we tested it among 5 users.They wear the proposed model,we calibrate the step length estimation FootStep-Tracker,walking along an approximately 100m- by efficiently employing the accelerometer combined with long path which is the full red line in Figure 3(b)in our de- gyroscopes. partment building.Users taked the elevator (the left bottom one)down to this floor,and moved along the path.After Step Direction Estimator.It measures the moving direc- he/she turned six times,he/she stopped at another elevator tion once we detect the turning steps from the heading in the map,and we stopped the tracking.Figure 3(c)shows forward steps.By merging consecutive turning steps in a the average location accuracy for each user.On average, whole turning process,and extracting the integral,max and the location accuracy is within 1m. variance as features,we classify the current turning direc- tions into three classes:left turn,right turn. Demo Application FootStep-Tracker interacts with user via an Android applica- Reference Position Estimator.It estimates the reference tion.Figure 4 shows the snapshot of the application.When positions of the moving trace based on the current activity user connecting the SensorTag and moving in the indoor and an indoor map.As the location of elevators and stairs environment,the application timely reports the users'step are fixed in the indoor environment,Reference Position Es- number,moving trace,current activity,sensors'data and timator treats the position of elevators or stairs as the refer- the user's location in the indoor map. ence position of the moving trace. Evaluation Acknowledgements This work is supported in part by National Natural Sci- Classification accuracy of Activity Classifier.To eval- ence Foundation of China under Grant Nos.61472185. uate the accuracy of the Activity Classifier,we embed the 61373129,61321491.91218302:Key Project of Jiangsu FootStep-Tracker into user's shoes and collect sensors'da- Research Program under Grant No.BE2013116;EU FP7 ta.The Sensor-Tag's sample frequency is set as 10 Hz,and IRSES MobileCloud Project under Grant No.612212.This it is embedded in the insole as depicted in Figure 1.For work is partially supported by Collaborative Innovation Cen- each activity,we collect about 500 windows of accelerom-Table 1: Label Description U_ST going upstairs D_ST going downstairs WALK walking E_HG hypergravity in elevator. E_WL weightlessness in elevator. E_UP elevator going up E_DOWN elevator going down SS stand still the proposed model, we calibrate the step length estimation by efficiently employing the accelerometer combined with gyroscopes. Step Direction Estimator. It measures the moving direc￾tion once we detect the turning steps from the heading forward steps. By merging consecutive turning steps in a whole turning process, and extracting the integral, max and variance as features, we classify the current turning direc￾tions into three classes: left turn, right turn. Reference Position Estimator. It estimates the reference positions of the moving trace based on the current activity and an indoor map. As the location of elevators and stairs are fixed in the indoor environment, Reference Position Es￾timator treats the position of elevators or stairs as the refer￾ence position of the moving trace. Evaluation Classification accuracy of Activity Classifier. To eval￾uate the accuracy of the Activity Classifier, we embed the FootStep-Tracker into user’s shoes and collect sensors’ da￾ta. The Sensor-Tag’s sample frequency is set as 10 Hz, and it is embedded in the insole as depicted in Figure 1. For each activity, we collect about 500 windows of accelerom￾eter and gyroscope data among 5 different users. Then we use them to train the classify model. We perform the clas￾sifier on the previous 5 users and 3 new users. They go up/downstairs, walk or take elevator. Figure 3 (a) shows the accuracy of Activity Classifier on the collected data. Particu￾larly, we don’t evaluate the EUP/EDOWN because that EUP is the combination of EHG and EWL, and EDOWN is the combination of EWL and EHG. On average, we achieve a classification accuracy of 96.2%. Location Accuracy. To evaluate the location performance of FootStep-Tracker, we tested it among 5 users. They wear FootStep-Tracker, walking along an approximately 100m￾long path which is the full red line in Figure 3 (b) in our de￾partment building. Users taked the elevator (the left bottom one) down to this floor, and moved along the path. After he/she turned six times, he/she stopped at another elevator in the map, and we stopped the tracking. Figure 3 (c) shows the average location accuracy for each user. On average, the location accuracy is within 1m. Demo Application FootStep-Tracker interacts with user via an Android applica￾tion. Figure 4 shows the snapshot of the application. When user connecting the SensorTag and moving in the indoor environment, the application timely reports the users’ step number, moving trace, current activity, sensors’ data and the user’s location in the indoor map. Acknowledgements This work is supported in part by National Natural Sci￾ence Foundation of China under Grant Nos. 61472185, 61373129, 61321491, 91218302; Key Project of Jiangsu Research Program under Grant No. BE2013116; EU FP7 IRSES MobileCloud Project under Grant No. 612212. This work is partially supported by Collaborative Innovation Cen-