2158 IEEE/ACM TRANSACTIONS ON NETWORKING,VOL.26.NO.5.OCTOBER 2018 000000C7 00.t40.0Qa020 DTE 0.0000000000040000000008 0000000 000000 Fh.c00c00c0000 0020000 n00 d00c0nd000000 Walk 0.00.040.000000000g A00c60.0000000000 0026ù20.0 (a) 00500 CCc00.12n.120000000000.00 o.0oo0 46000000016000000Q03Q3g9.0g Walk Run Rest mp Walk P970000020000000020.000500.0时 B92.c000000000000a.00a03Q.03 Bs2.800.000u000000000.00000003 DTE DLAUBR CC BF DF RS DC.PS BS DTE DLRUBR CC BF DF RS DC PS BS (a) (b) OTE0■ c00C0022000000000003 DTE 0000c00 000080000050050.00 00G0002000002003 D月象80I000002uw0g0002002003Q0 UB时h.00.00 10000D.10D.0000 UE明3.800000000ND p0340.03014 cc.000.020.0 00000000040.0g 2009Q04Q10 F000.0000.0 004 904200 mc0a00n000000m 05Q0 07 3600800800800G0 0F力80G000000000003000 Sampling Sequence (50Hz) 2000 500 000000800000C0000. 16000400000000000000000 (b) 0000.1202000000000 0 0000600200000003 40004000000024 00 0000200 00003.0 Fig.17.Experiment results.(a)An example process of human movement (b)The variation of angle/coordinate accuracy. (d) 93 Fig.19.Activity recognition accuracy.(a)DTW-based Recognition based on Compass sychronization.(b)DTW-based Recognition based on MOSS sychronization.(c)RF-based Recognition based on Compass sychronization. (d)RF-based Recognition based on MOSS sychronization. the average recognition accuracy is 84.2%for the MOSS De synchronized coordinate.In regard to the Random For- (a) b est method,the average recognition accuracy is 85.6%for the Compass synchronized coordinate,whereas the average recognition accuracy is 73.2%for the MOSS synchronized coordinate.This implies that the recognition performance is not dramatically degraded based on MOSS synchronization, since MOSS achieves fairly good performance in space syn- @ (d) chronization Fig.18.Accuracy evaluation.(a)Angle accuracy of CFDE.(b)Coordinate accuracy of CFDE.(c)Angle accuracy of GOT.(d)Coordinate accuracy of XI.CONCLUSION GOT. We made three key contributions in this paper.First, we investigate the problem of space synchronization for than 87%,while the average coordinate accuracy is 91%.For mobile devices.Second,we propose the MOSS scheme to the performance of GOT,as the human subject stops moving achieve space synchronization among multiple mobile devices. forward,we evaluated the accuracy after a time interval In particular,we propose a consistent direction estimator of 30 seconds.As shown in Fig.18(c),the angle accuracies to achieve space synchronization,and a gyroscope based of all devices are less than 26,while the average accuracy orientation tracker to maintain space synchronization.Third, is 21,even if the time interval is as long as 30s.In Fig.18(d), we implemented MOSS on COTS mobile devices.and the the coordinate accuracies of all devices are greater than 80%, experiment results show that MOSS achieves an average while the average coordinate accuracy is 84%. angle accuracy of 9.8 and an average coordinate accuracy of97%. B.Activity Recognition Accuracy REFERENCES Based on the synchronized coordinates,we further eval- [1]S.Shen,H.Wang,and R.R.Choudhury,"I am a smartwatch and i can uate the performance in activity recognition,by using the track my user's arm."in Proc.Mobisys,2016.pp.85-96. methods of Dynamic Time Warping (DTW)and Random [2]L.Zhang et al.,"It starts with iGaze:Visual attention driven networking Forest(RF).Specifically,we let 10 volunteers perform 10 cat- with smart glasses,"in Proc.MobiCom,2014,pp.91-102. [3]C.Karatas et al,"Leveraging wearables for steering and driver track- egories of activities,including Dumbbell Triceps Exten- ing."in Proc.IEEE INFOCOM,Apr.2016.pp.1-9. sion(DTE),Dumbbell Lateral Raise (DLR),Upright Barbell [4]J.Yu et al.,"SenSpeed:Sensing driving conditions to estimate vehicle Row (UBR),Cable Crossover(CC),Butterfly (BF),Dumbbell speed in urban environments,"IEEE Trans.Mobile Comput.,vol.15. n0.1,Pp.202-216Jan.2016. Flies(DF),Rope Skipping(RS),Dumbbell Curl (DC),Ping- [5]Z.Sun et al.,"Polaris:Getting accurate indoor orientations for mobile pong Swing (PS),and Badminton Swing (BS).In order devices using ubiquitous visual pattems on ceilings,"in Proc.ACM HotMobile,2012,pp.14:1-14:6. to evaluate the impact on the recognition accuracy from [6]P.Zhou,M.Li,and G.Shen,"Use it free:Instantly knowing your phone the space synchronization,we plot the confusion matrix of attitude,"in Proc.MobiCom,2014,pp.605-616. activity recognition,respectively,based on Compass syn- [7]S.Poddar,V.Kumar,and A.Kumar,"A comprehensive overview of inertial sensor calibration techniques,"Dyn.Syst.,Meas.Control, chronized coordinates and MOSS synchronized coordinates. voL.139,no.1,p.011006.2016. Fig.19 shows the experiment results.Note that in regard [8]H.Fourati,"Heterogeneous data fusion algorithm for pedestrian navi- to the DTW method,the average recognition accuracy is gation via foot-mounted inertial measurement unit and complementary filter,"IEEE Trans.Instrum.Meas.,vol.64,no.1,pp.221-229. 84.6%for the Compass synchronized coordinate,whereas Jan.2015.2158 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 26, NO. 5, OCTOBER 2018 Fig. 17. Experiment results. (a) An example process of human movement. (b) The variation of angle/coordinate accuracy. Fig. 18. Accuracy evaluation. (a) Angle accuracy of CFDE. (b) Coordinate accuracy of CFDE. (c) Angle accuracy of GOT. (d) Coordinate accuracy of GOT. than 87%, while the average coordinate accuracy is 91%. For the performance of GOT, as the human subject stops moving forward, we evaluated the accuracy after a time interval of 30 seconds. As shown in Fig. 18(c), the angle accuracies of all devices are less than 26◦, while the average accuracy is 21◦, even if the time interval is as long as 30s. In Fig. 18(d), the coordinate accuracies of all devices are greater than 80%, while the average coordinate accuracy is 84%. B. Activity Recognition Accuracy Based on the synchronized coordinates, we further eval￾uate the performance in activity recognition, by using the methods of Dynamic Time Warping (DTW) and Random Forest (RF). Specifically, we let 10 volunteers perform 10 cat￾egories of activities, including Dumbbell Triceps Exten￾sion (DTE), Dumbbell Lateral Raise (DLR), Upright Barbell Row (UBR), Cable Crossover (CC), Butterfly (BF), Dumbbell Flies (DF), Rope Skipping (RS), Dumbbell Curl (DC), Ping￾pong Swing (PS), and Badminton Swing (BS). In order to evaluate the impact on the recognition accuracy from the space synchronization, we plot the confusion matrix of activity recognition, respectively, based on Compass syn￾chronized coordinates and MOSS synchronized coordinates. Fig. 19 shows the experiment results. Note that in regard to the DTW method, the average recognition accuracy is 84.6% for the Compass synchronized coordinate, whereas Fig. 19. Activity recognition accuracy. (a) DTW-based Recognition based on Compass sychronization. (b) DTW-based Recognition based on MOSS sychronization. (c) RF-based Recognition based on Compass sychronization. (d) RF-based Recognition based on MOSS sychronization. the average recognition accuracy is 84.2% for the MOSS synchronized coordinate. In regard to the Random For￾est method, the average recognition accuracy is 85.6% for the Compass synchronized coordinate, whereas the average recognition accuracy is 73.2% for the MOSS synchronized coordinate. This implies that the recognition performance is not dramatically degraded based on MOSS synchronization, since MOSS achieves fairly good performance in space syn￾chronization. XI. CONCLUSION We made three key contributions in this paper. First, we investigate the problem of space synchronization for mobile devices. Second, we propose the MOSS scheme to achieve space synchronization among multiple mobile devices. In particular, we propose a consistent direction estimator to achieve space synchronization, and a gyroscope based orientation tracker to maintain space synchronization. Third, we implemented MOSS on COTS mobile devices, and the experiment results show that MOSS achieves an average angle accuracy of 9.8◦ and an average coordinate accuracy of 97%. REFERENCES [1] S. Shen, H. Wang, and R. R. Choudhury, “I am a smartwatch and i can track my user’s arm,” in Proc. MobiSys, 2016, pp. 85–96. [2] L. Zhang et al., “It starts with iGaze: Visual attention driven networking with smart glasses,” in Proc. MobiCom, 2014, pp. 91–102. [3] C. Karatas et al., “Leveraging wearables for steering and driver track￾ing,” in Proc. IEEE INFOCOM, Apr. 2016, pp. 1–9. [4] J. Yu et al., “SenSpeed: Sensing driving conditions to estimate vehicle speed in urban environments,” IEEE Trans. Mobile Comput., vol. 15, no. 1, pp. 202–216, Jan. 2016. [5] Z. Sun et al., “Polaris: Getting accurate indoor orientations for mobile devices using ubiquitous visual patterns on ceilings,” in Proc. ACM HotMobile, 2012, pp. 14:1–14:6. [6] P. Zhou, M. Li, and G. Shen, “Use it free: Instantly knowing your phone attitude,” in Proc. MobiCom, 2014, pp. 605–616. [7] S. Poddar, V. Kumar, and A. Kumar, “A comprehensive overview of inertial sensor calibration techniques,” J. Dyn. Syst., Meas. Control, vol. 139, no. 1, p. 011006, 2016. [8] H. Fourati, “Heterogeneous data fusion algorithm for pedestrian navi￾gation via foot-mounted inertial measurement unit and complementary filter,” IEEE Trans. Instrum. Meas., vol. 64, no. 1, pp. 221–229, Jan. 2015