AI A2 A3 AI A2A3 Moving T Tag receiving antennas are attached on each wall randomly.We deploy the transmitting antenna in the center of room for Stationary Updated simplicity.The phase values are measured according to the Taes Distribution transmitting distance and wavelength.We first extract the Stationary Distribution stationary phase distribution from singleton slot in the tag Fig.11:Matching example inventory phase.Then in continuous polling phase,we extract the real-time phase distribution from singleton and collision VII.PERFORMANCE EVALUATION slots.For every polling cycle,we utilize the Hungarian algo- In the following,we evaluate the performance from three rithm to detect the moving tags. aspects.Firstly,we evaluate the feature extraction scheme In this experiment,we study the effect of the following based on a small scale experiment in the realistic environment parameters on the detection accuracy:phase variance o,num- Then we evaluate GMM solution in terms of the detection ber of moving tags n,number of antennas r and number of accuracy of motion status.Lastly,we compare FTPS and monitoring tagsw.K is set to 2w for simplicity.We utilize false C1G2 standard to evaluate the time efficiency. positive error(FP)and false negative error(FN)to evaluate the accuracy.False positive error means the stationary tags is A.Evaluate the Accuracy in Feature Extraction identified as moving tags,and false negative error means the 1)Experiment Settings:We perform a realistic experiment moving tags is identified as stationary tags. to evaluate the extracted physical-layer features by issuing 50 Because phase values vary in accordance with Gaussian tags.All signal traces are collected with the GNURadio/USRP distribution as evaluated.We import the variances according to platform as described in Section IV.Due to the power limita- the collision type.For singleton slot,a fixed standard deviation tion of USRP for scattering 50 tags,we use a four-step scheme of 4 is used based on the evaluation in Section IV-B.For to collect the realistic signal.Firstly,we emulate the process collision slot,we import parameters o and 2o to denote the of slot selection on MATLAB according to C1G2 standard. standard deviations of 2-collision and 3-collision respectively. Secondly,we collect the responding signal trace for each slot We cannot extracted phase profiles from 4+-collision slots. on the USRP platform according to the result of slot selection. 2)Results:GMM can correctly detect 85%of the moving Thirdly,we extract physical-layer features from singleton and tags with about 6%false positive errors.Fig.12(e)~12(h) collision slots.Lastly,we modify the frame size based on the report the results from different points of views.We note the number of identified tags for a new frame.In this experiment, number of antennas and tags affect the accuracy obviously, the initial frame size is set to 64 according the tag cardinality. while the number of moving tags slightly affect the accuracy. 2)Results:We can save 60%frames to extract physical- Specifically,FN errors are mainly caused by the relative layer features by decoding the collision signals,while the scales between antenna number and tag number.When the standard variance of extracted phase profiles and BLF are number of antennas is relatively small compared with the 9.7 and 4.3us respectively.Fig.12(a)shows the collision number of tags,we cannot exclusively distinguish all tags distribution and corresponding slot utilization of the first due to lower dimension of phase profiles.As a result,some frame.When we extract features from singleton,2-collision moving tags cannot be detected,which leads to high FN errors. and 3-collision slots,there are only 4.6 tags unidentified on As the number of antennas increases to be comparable with average after the first frame.The identification ratio is about the cardinality of tags,we can accurately detect the moving 2 times of C1G2 standard,which only identifies tags in the tags with high probability.FP errors are mainly caused by singleton slot.Further,Fig.12(b)reports unidentified tags the measuring phase variance.When the phase profiles are number of each frame.To identify all the tags,ClG2 standard relatively accurate,it is less likely to identify a stationary tag as usually need 5 frames typically.However,if we could exploit a moving one.Besides,both FP and FN are not sensitive about the collision slot,we could finish the inventory in 2 frames. the number of moving tags.This is because when fixing other Meanwhile,as shown in Fig.12(c)and 12(d),both phase parameters,the probability of correct and incorrect identifying profiles and BLF variance can be sufficiently restricted.We each tag is the same.Hence in Fig.12(g)both FP and FN exploit the cumulative distribution function(CDF)to show the varies a little. accuracy of the extracted physical-layer features from collision C.Evaluate the Time Efficiency signal.80%variances of signal length are 5us and 8us for 2- FTPS can save 80%inventory time when querying more collision and 3-collision respectively.For the phase profile, than hundreds of tags compared with CIG2 protocol.We the variances of 80%sample are 8 and 16 respectively.The compare with C1G2 standard to evaluate the time efficiency. errors are caused by the noise imported by the collided tags. Specifically,we evaluate the querying time and the cost frames of conducting a typical query cycle.We use the time of each B.Evaluate the Accuracy in Moving Tag Detection period in [1,2]to estimate the querying time,i.e.,Ims for 1)Experiment Settings:We further perform extensive sim- collision slot and 4ms for singleton slot.We set the initial ulations to evaluate GMM over different parameters on MAT- frame size based on the number of tags. LAB.In this experiment,we monitor hundreds of tags,which As shown in Fig.13(a),the modified protocol can save are scattered randomly inside a 10m x 10m room and the 80%time compared with C1G2 standard through decoding the!"#$%&' ()& *+)+$"%),-' ()&. /0 /1 /2 /0 /1 /2 345)+65' 7$.+,$89+$"% *+)+$"%),-' 7$.+,$89+$"% ! !" Fig. 11: Matching example VII. PERFORMANCE EVALUATION In the following, we evaluate the performance from three aspects. Firstly, we evaluate the feature extraction scheme based on a small scale experiment in the realistic environment. Then we evaluate GMM solution in terms of the detection accuracy of motion status. Lastly, we compare FTPS and C1G2 standard to evaluate the time efficiency. A. Evaluate the Accuracy in Feature Extraction 1) Experiment Settings: We perform a realistic experiment to evaluate the extracted physical-layer features by issuing 50 tags. All signal traces are collected with the GNURadio/USRP platform as described in Section IV. Due to the power limita￾tion of USRP for scattering 50 tags, we use a four-step scheme to collect the realistic signal. Firstly, we emulate the process of slot selection on MATLAB according to C1G2 standard. Secondly, we collect the responding signal trace for each slot on the USRP platform according to the result of slot selection. Thirdly, we extract physical-layer features from singleton and collision slots. Lastly, we modify the frame size based on the number of identified tags for a new frame. In this experiment, the initial frame size is set to 64 according the tag cardinality. 2) Results: We can save 60% frames to extract physical￾layer features by decoding the collision signals, while the standard variance of extracted phase profiles and BLF are 9.7◦ and 4.3µs respectively. Fig. 12(a) shows the collision distribution and corresponding slot utilization of the first frame. When we extract features from singleton, 2-collision and 3-collision slots, there are only 4.6 tags unidentified on average after the first frame. The identification ratio is about 2 times of C1G2 standard, which only identifies tags in the singleton slot. Further, Fig. 12(b) reports unidentified tags number of each frame. To identify all the tags, C1G2 standard usually need 5 frames typically. However, if we could exploit the collision slot, we could finish the inventory in 2 frames. Meanwhile, as shown in Fig. 12(c) and 12(d), both phase profiles and BLF variance can be sufficiently restricted. We exploit the cumulative distribution function (CDF) to show the accuracy of the extracted physical-layer features from collision signal. 80% variances of signal length are 5µs and 8µs for 2- collision and 3-collision respectively. For the phase profile, the variances of 80% sample are 8◦ and 16◦ respectively. The errors are caused by the noise imported by the collided tags. B. Evaluate the Accuracy in Moving Tag Detection 1) Experiment Settings: We further perform extensive sim￾ulations to evaluate GMM over different parameters on MAT￾LAB. In this experiment, we monitor hundreds of tags, which are scattered randomly inside a 10m × 10m room and the receiving antennas are attached on each wall randomly. We deploy the transmitting antenna in the center of room for simplicity. The phase values are measured according to the transmitting distance and wavelength. We first extract the stationary phase distribution from singleton slot in the tag inventory phase. Then in continuous polling phase, we extract the real-time phase distribution from singleton and collision slots. For every polling cycle, we utilize the Hungarian algo￾rithm to detect the moving tags. In this experiment, we study the effect of the following parameters on the detection accuracy: phase variance σ, num￾ber of moving tags η, number of antennas τ and number of monitoring tags ω. κ is set to 2ω for simplicity. We utilize false positive error (FP) and false negative error (FN) to evaluate the accuracy. False positive error means the stationary tags is identified as moving tags, and false negative error means the moving tags is identified as stationary tags. Because phase values vary in accordance with Gaussian distribution as evaluated. We import the variances according to the collision type. For singleton slot, a fixed standard deviation of 4◦ is used based on the evaluation in Section IV-B. For collision slot, we import parameters σ and 2σ to denote the standard deviations of 2-collision and 3-collision respectively. We cannot extracted phase profiles from 4+-collision slots. 2) Results: GMM can correctly detect 85% of the moving tags with about 6% false positive errors. Fig. 12(e)∼12(h) report the results from different points of views. We note the number of antennas and tags affect the accuracy obviously, while the number of moving tags slightly affect the accuracy. Specifically, FN errors are mainly caused by the relative scales between antenna number and tag number. When the number of antennas is relatively small compared with the number of tags, we cannot exclusively distinguish all tags due to lower dimension of phase profiles. As a result, some moving tags cannot be detected, which leads to high FN errors. As the number of antennas increases to be comparable with the cardinality of tags, we can accurately detect the moving tags with high probability. FP errors are mainly caused by the measuring phase variance. When the phase profiles are relatively accurate, it is less likely to identify a stationary tag as a moving one. Besides, both FP and FN are not sensitive about the number of moving tags. This is because when fixing other parameters, the probability of correct and incorrect identifying each tag is the same. Hence in Fig. 12(g) both FP and FN varies a little. C. Evaluate the Time Efficiency FTPS can save 80% inventory time when querying more than hundreds of tags compared with C1G2 protocol. We compare with C1G2 standard to evaluate the time efficiency. Specifically, we evaluate the querying time and the cost frames of conducting a typical query cycle. We use the time of each period in [1, 2] to estimate the querying time, i.e., 1ms for collision slot and 4ms for singleton slot. We set the initial frame size based on the number of tags. As shown in Fig. 13(a), the modified protocol can save 80% time compared with C1G2 standard through decoding the