IEEE TRANSACTIONS ON MOBILE COMPUTING,VOL.XX,NO.XX,2019 11 5.2.1 Orientation Accuracy with Different Scanning Ranges with or without Outliers Elimination ntenna Tags Tagged Boxes. The most advantage of our approach compared to STPP [19] is that we do not require the large range scanning.We put boxl in front of the antenna plane about 0.8m,and adjust Reader scanning ranges from 0.3m to 0.9m on a single side or both Moving Car sides.Taking range of 0.3m for example,for a single side,it anning means the antenna starts scanning from the box and moves 0.3m.While for both sides,the scanning range is 0.6m, Fig.19.Deployment of RF-3DScan centered on the box.To validate the efficiency of outliers elimination during the linear fitting process,we first use a practical warehouse,there are several sizes of packages, all data to do the linear fitting,then eliminate outliers for it is possible that several small packages are stacked on a comparison.The results without the outliers elimination are big package.If we still view packages as points,we cannot shown in Fig.20-21.From Fig.20,RF-3DScan achieves the determine such complex stacking.Therefore,to handle with average bottom face accuracies about 95%for both sides the complex situation,after determining the orientation of scanning,and about 70%for one side scanning range of the tag array,it is necessary to reconstruct the package in the 0.7m.When the both scanning range achieves 0.5m,RF- 3D space based on the priori knowledge,i.e.,the geometric 3DScan can identify the bottom face totally correctly.For relationship of the tag array attached on the package.Then, the single side scanning,RF-3DScan fails to work with the we can determine the complex stacking by comparing the scanning range less than 0.5m.With the range extending projected polygons of packages in different planes.Specif- to 0.7m,the accuracy increases to 70%.From Fig.21,the ically,by comparing the relationship of projected polygons average angle error for the both sides scanning is smaller of packages in the XY plane,i.e.,parallel to the ground, than that for the single side scanning.Also,the angle error we can determine which packages are likely to be stacked will not be affected by the scanning range significantly. together.Then,by comparing the relationship of projected Though the accuracy of the one side scanning is not that polygons of packages in the XZ and YZ plane,we can good,STPP cannot even deal with such limited scanning determine the left-right and up-down relationship among ranges.The similar results with the outliers elimination packages in a complex stack. are shown in Fig.22-23.With the outliers elimination,the obvious improvement is obtained when the scanning range is larger than 0.7m. 6 PERFORMANCE EVALUATION 5.1 Experiment Settings 5.2.2 Orientation Accuracy with Different Tag Separation Distances We build a prototype system of RF-3DScan,as shown in Fig.19.Hardware:Our system consists of one Impin] We adjust the tag separation of box3 from 11cm to 23cm, Speedway R420 reader,one Laird S9028 RFID antenna and which is in front of the antenna plane about 1m,and let multiple ImpinJ E41-B tags.The antenna is fixed on a mov- the antenna perform both sides scanning of 0.5m 40 times ing car with the height of about 1.8m.Software:We adopt for each separation.As shown in Fig.24,with the larger LLRP protocol to communicate with the reader,and use tag separation,the bottom face accuracy gets higher and a special module to control the movement of the car.Our the angle error gets smaller correspondingly.It is found algorithms are implemented in MATLAB and the language that when the tag separation is no less than 17cm,the performance tends to be stable,and we can identify the Java.Deployment:We let the antenna be above the boxes and move at a constant speed of 0.12m/s.For diversity, bottom face of the box accurately with the angle error we use three different sizes of boxes.The sizes of box1, smaller than 3.05.While when the tag separation is smaller box2 and box.3are0.4m×0.35m×0.33m,0.35m×0.27×0.4m, than 17cm,the accuracy is seriously affected by the change 0.49mx0.39mx0.32m,respectively.For each box,there are of tag separation.According to previous work Tagyro [6], six tags on it as shown in Fig.10(c),the tag separation of when the tag separation is larger than 15cm,the coupling two tags on the same surface is the same,as 0.23m,0.17m effect between tags is relatively small,while for the smaller and 0.2cm for box1,box2 and box3,respectively. tag separation,the coupling effect gets stronger very fast, which agrees with our experiment results.The coupling effect can be handled with the method proposed by Tagyro, 5.2 Micro-Benchmarks which is beyond the scope of this paper,so we choose the Metrics:To evaluate the package orientation accuracy,we set tag separation larger than 17cm to limit the coupling effect. two metrics:bottom face accuracy and angle error.The bottom face accuracy is defined as the number of the packages 5.2.3 Orientation Accuracy with Different Contents whose bottom faces are identified correctly out of the total In order to evaluate the effect of items made of different ma- package number.The angle error is the error between the terials in the box,we fill box3 with items made of different estimated angle of the vertical faces against the antenna materials,like plastic toys (Pla),cotton pants (Cot),foam plane and the actual angle.For the package stacking,we (Foa),mineral water (Liq),antennas made of plastic and use the metric ordering accuracy.The box is ordered correctly metal(Mix),or just let the box empty (Emp)for comparison. only when its detected order is same as the actual order. We perform the scanning with the same settings as above.IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. XX, NO. XX, 2019 11 1st Linear Scanning 2nd Linear Scanning Antenna Reader Moving Car Tagged Boxes Tags Fig. 19. Deployment of RF-3DScan a practical warehouse, there are several sizes of packages, it is possible that several small packages are stacked on a big package. If we still view packages as points, we cannot determine such complex stacking. Therefore, to handle with the complex situation, after determining the orientation of the tag array, it is necessary to reconstruct the package in the 3D space based on the priori knowledge, i.e., the geometric relationship of the tag array attached on the package. Then, we can determine the complex stacking by comparing the projected polygons of packages in different planes. Specif￾ically, by comparing the relationship of projected polygons of packages in the XY plane, i.e., parallel to the ground, we can determine which packages are likely to be stacked together. Then, by comparing the relationship of projected polygons of packages in the XZ and Y Z plane, we can determine the left-right and up-down relationship among packages in a complex stack. 5 PERFORMANCE EVALUATION 5.1 Experiment Settings We build a prototype system of RF-3DScan, as shown in Fig.19. Hardware: Our system consists of one ImpinJ Speedway R420 reader, one Laird S9028 RFID antenna and multiple ImpinJ E41-B tags. The antenna is fixed on a mov￾ing car with the height of about 1.8m. Software: We adopt LLRP protocol to communicate with the reader, and use a special module to control the movement of the car. Our algorithms are implemented in MATLAB and the language Java. Deployment: We let the antenna be above the boxes and move at a constant speed of 0.12m/s. For diversity, we use three different sizes of boxes. The sizes of box1, box2 and box3 are 0.4m×0.35m×0.33m, 0.35m×0.27×0.4m, 0.49m×0.39m×0.32m, respectively. For each box, there are six tags on it as shown in Fig. 10(c), the tag separation of two tags on the same surface is the same, as 0.23m, 0.17m and 0.2cm for box1, box2 and box3, respectively. 5.2 Micro-Benchmarks Metrics: To evaluate the package orientation accuracy, we set two metrics: bottom face accuracy and angle error. The bottom face accuracy is defined as the number of the packages whose bottom faces are identified correctly out of the total package number. The angle error is the error between the estimated angle of the vertical faces against the antenna plane and the actual angle. For the package stacking, we use the metric ordering accuracy. The box is ordered correctly only when its detected order is same as the actual order. 5.2.1 Orientation Accuracy with Different Scanning Ranges with or without Outliers Elimination The most advantage of our approach compared to STPP [19] is that we do not require the large range scanning. We put box1 in front of the antenna plane about 0.8m, and adjust scanning ranges from 0.3m to 0.9m on a single side or both sides. Taking range of 0.3m for example, for a single side, it means the antenna starts scanning from the box and moves 0.3m. While for both sides, the scanning range is 0.6m, centered on the box. To validate the efficiency of outliers elimination during the linear fitting process, we first use all data to do the linear fitting, then eliminate outliers for comparison. The results without the outliers elimination are shown in Fig. 20-21. From Fig. 20, RF-3DScan achieves the average bottom face accuracies about 95% for both sides scanning, and about 70% for one side scanning range of 0.7m. When the both scanning range achieves 0.5m, RF- 3DScan can identify the bottom face totally correctly. For the single side scanning, RF-3DScan fails to work with the scanning range less than 0.5m. With the range extending to 0.7m, the accuracy increases to 70%. From Fig. 21, the average angle error for the both sides scanning is smaller than that for the single side scanning. Also, the angle error will not be affected by the scanning range significantly. Though the accuracy of the one side scanning is not that good, STPP cannot even deal with such limited scanning ranges. The similar results with the outliers elimination are shown in Fig. 22-23. With the outliers elimination, the obvious improvement is obtained when the scanning range is larger than 0.7m. 5.2.2 Orientation Accuracy with Different Tag Separation Distances We adjust the tag separation of box3 from 11cm to 23cm, which is in front of the antenna plane about 1m, and let the antenna perform both sides scanning of 0.5m 40 times for each separation. As shown in Fig. 24, with the larger tag separation, the bottom face accuracy gets higher and the angle error gets smaller correspondingly. It is found that when the tag separation is no less than 17cm, the performance tends to be stable, and we can identify the bottom face of the box accurately with the angle error smaller than 3.05◦ . While when the tag separation is smaller than 17cm, the accuracy is seriously affected by the change of tag separation. According to previous work Tagyro [6], when the tag separation is larger than 15cm, the coupling effect between tags is relatively small, while for the smaller tag separation, the coupling effect gets stronger very fast, which agrees with our experiment results. The coupling effect can be handled with the method proposed by Tagyro, which is beyond the scope of this paper, so we choose the tag separation larger than 17cm to limit the coupling effect. 5.2.3 Orientation Accuracy with Different Contents In order to evaluate the effect of items made of different ma￾terials in the box, we fill box3 with items made of different materials, like plastic toys (Pla), cotton pants (Cot), foam (Foa), mineral water (Liq), antennas made of plastic and metal (Mix), or just let the box empty (Emp) for comparison. We perform the scanning with the same settings as above