2156 IEEE/ACM TRANSACTIONS ON NETWORKING,VOL.26.NO.5.OCTOBER 2018 到喜型D五0☑ 多☒四四 (a b (c) (d) (c) Fig.14.Evaluate the accuracy of consistent forwarding direction estimation.(a)Angle accuracy for different forwarding direction estimators.(b)Angle accuracy for different moving modes.(c)Angle accuracy with different number of devices.(d)Coordinate accuracy for different forwarding direction estimators. (e)Coordinate accuracy for different moving modes.(f)Coordinate accuracy with different number of devices. ope Tracking Reference Comp-Flter Horizonta Dresson (a) (b) (c) Fig.15.Evaluate the accuracy of gyroscope-based orientation tracking.(a)Error accumulation over time for different solutions.(b)Coordinate accuracy for orientation tracking after different time intervals.(c)Coordinate accuracy for orientation tracking in different moving modes to sudden,rapid and irregular acceleration and orientation coordinate and ground truth.For MOSS,the average correla- changes,the Kalman filter-based solution can effectively filter tions from six devices are all greater than 95%.For MeanACC, the corresponding inertial readings such that the performance most of the similarities are less than 60%,the similarity in S cannot be degraded too much. is even as low as 49%.For RefMeanACC,the performance Experimental results show that MOSS achieves average is increased to some extent,however,the average similarity is angle accuracy of 9.8 for all devices in different moving only 73%.We further evaluated the correlations of measure- modes and with different time windows.We first evaluated ments in different moving modes as shown in Fig.14(e). MOSS when the human subject is walking.Fig.14(a)plots MOSS achieves good performance in measurement correla- the angle deviation between the estimated forwarding direction tions for all three modes. and the ground truth,where we show the mean and standard Experimental results show that,when the number of devices deviation for all six devices.For MOSS,the average angle is varied from 3~6,MOSS achieved fairly good performance deviation in six devices are all less than 15 while the standard in average angle accuracy and average coordinate accuracy. deviations are usually less than 9.For both MeanACC and We vary the number of devices from 3 to 6,and evaluate RefMeanACC,the average angle deviations are much greater the angle accuracy and coordinate accuracy,respectively.As than MOSS,besides,they all have much greater variances shown in Fig.14(c)and Fig.14(f),it is found that,when the than MOSS.We further evaluated the average angle deviation number of devices is decreased from 6 to 3,both the angle respectively in walk,run,and jump,as shown in Fig.14(b).accuracy and coordinate accuracy is decreased accordingly. We observed that MOSS achieves fairly good performance in e.g.,when the number of devices is decreased from 6 to 3, angle accuracy for all three modes,which implies that MOSS the average angle deviation is increased from 9.8 to 33 and is not very sensitive to the exact moving mode.Specifically.the average similarity is is decreased from to 97%to 70%. MOSS achieves best performance in the jump mode,as it This is caused by the reduced number of streams in the PCA generates larger consistent forwarding accelerations to assist process.Nevertheless,it still achieves fairly good performance space synchronization.The performance in the run mode on the whole. degrades to some extent,as it generates larger inconsistent accelerations due to larger movements of the limbs. C.Gyroscope-Based Orientation Tracking Experimental results show that MOSS achieved average We implemented the following solutions for performance coordinate accuracy of 97%for all devices in different moving comparison:1)Gyroscope Tracking:the orientation tracking modes and with different time windows.We first evaluated the scheme purely based on the gyroscope measurements [12]. performance when the human subject is walking.Fig.14(d)2)Reference Complementary Filter:a common complemen- plots the correlations of measurements from the synchronized tary filter combining the accelerometer and gyroscope data2156 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 26, NO. 5, OCTOBER 2018 Fig. 14. Evaluate the accuracy of consistent forwarding direction estimation. (a) Angle accuracy for different forwarding direction estimators. (b) Angle accuracy for different moving modes. (c) Angle accuracy with different number of devices. (d) Coordinate accuracy for different forwarding direction estimators. (e) Coordinate accuracy for different moving modes. (f) Coordinate accuracy with different number of devices. Fig. 15. Evaluate the accuracy of gyroscope-based orientation tracking. (a) Error accumulation over time for different solutions. (b) Coordinate accuracy for orientation tracking after different time intervals. (c) Coordinate accuracy for orientation tracking in different moving modes. to sudden, rapid and irregular acceleration and orientation changes, the Kalman filter-based solution can effectively filter the corresponding inertial readings such that the performance cannot be degraded too much. Experimental results show that MOSS achieves average angle accuracy of 9.8◦ for all devices in different moving modes and with different time windows. We first evaluated MOSS when the human subject is walking. Fig. 14(a) plots the angle deviation between the estimated forwarding direction and the ground truth, where we show the mean and standard deviation for all six devices. For MOSS, the average angle deviation in six devices are all less than 15◦ while the standard deviations are usually less than 9◦. For both MeanACC and RefMeanACC, the average angle deviations are much greater than MOSS, besides, they all have much greater variances than MOSS. We further evaluated the average angle deviation respectively in walk, run, and jump, as shown in Fig. 14(b). We observed that MOSS achieves fairly good performance in angle accuracy for all three modes, which implies that MOSS is not very sensitive to the exact moving mode. Specifically, MOSS achieves best performance in the jump mode, as it generates larger consistent forwarding accelerations to assist space synchronization. The performance in the run mode degrades to some extent, as it generates larger inconsistent accelerations due to larger movements of the limbs. Experimental results show that MOSS achieved average coordinate accuracy of 97% for all devices in different moving modes and with different time windows. We first evaluated the performance when the human subject is walking. Fig. 14(d) plots the correlations of measurements from the synchronized coordinate and ground truth. For MOSS, the average correla￾tions from six devices are all greater than 95%. For MeanACC, most of the similarities are less than 60%, the similarity in S4 is even as low as 49%. For RefMeanACC, the performance is increased to some extent, however, the average similarity is only 73%. We further evaluated the correlations of measure￾ments in different moving modes as shown in Fig. 14(e). MOSS achieves good performance in measurement correla￾tions for all three modes. Experimental results show that, when the number of devices is varied from 3∼6, MOSS achieved fairly good performance in average angle accuracy and average coordinate accuracy. We vary the number of devices from 3 to 6, and evaluate the angle accuracy and coordinate accuracy, respectively. As shown in Fig. 14(c) and Fig. 14(f), it is found that, when the number of devices is decreased from 6 to 3, both the angle accuracy and coordinate accuracy is decreased accordingly, e.g., when the number of devices is decreased from 6 to 3, the average angle deviation is increased from 9.8◦ to 33◦ and the average similarity is is decreased from to 97% to 70%. This is caused by the reduced number of streams in the PCA process. Nevertheless, it still achieves fairly good performance on the whole. C. Gyroscope-Based Orientation Tracking We implemented the following solutions for performance comparison: 1) Gyroscope Tracking: the orientation tracking scheme purely based on the gyroscope measurements [12]. 2) Reference Complementary Filter: a common complemen￾tary filter combining the accelerometer and gyroscope data