EEE ICC 2014-Ad-hoc and Sensor Networking Symposium is an intermediate result in algorithm 1.By setting the power to po and scan the tags,we get the identified categories Ci and the RSSI value si.j of each tag in the category Ci.Through 1111101 the training data set Ta,we can get the distance di.j of the tag to the center of the detection region.Take the position of the category's midpoint as the tag's position,then the estimation position i calculated by the category Ci is .The category contains the minimal dij is the most possible category which contains the real position Fig.8.Experiments setup xo,we can ensure the relative positions of the categories.Then the estimation position is power used by the baseline protocols.Usually,the larger power means the larger detected tag size,which leads to more Algorithm 3 Distance Voting based protocol execution time.In Fig.9,the tag density is 10tags/m and the 1:Call Alg.I to compute the tag density p and the optimal average width of categories is 0.5m,the execution time of value of power po. CMP is only 62%of that of CCP,while the execution time of 2:Get the identified categories Ci=fi.s,i,e,li}and the DVP is only of 57%of that of RP.In regard to the proposed RSSI value si.i of each tag in the category Ci. protocols,DVP has larger execution time than CMP,because 3:Through the training data T,get the distance di.j of the DVP needs to get the tag IDs and the RSSI value of each tags tag to the center of the detection region. as well. 4:Find the category Co which has the minimal di.j. 5:fori∈{1,2k}j∈{1,2ni}do C.Accuracy of Localization 6: if>othen d endif From Fig.10,we can find that all our proposed protocols have higher accuracies than the baseline protocol.Besides, 7: io then endif 2 the proposed protocols are more reliable than the baseline 8:end for 9:compute the交as follows::金=∑t,色 protocols with lower localization error.This is because our proposed protocol(CMP,DVP)respectively use the category 10:OUTPUT:the estimation position. matching method and the distance voting method to select the k nearest categories to locate the target.They reduce the error 3)Analysis:This method improves the reliablity of the caused by the variances of tag distributions.When the tag baseline RSSI weighted average based method.By getting density is 10tags/m and the average category width is 0.5m, the corresponding distance value of the RSSI,we make the the localization error of CMP is only 75%of that of CCP, algorithm more robust. while the localization error of DVP is only 73%of that of RP. VII.PERFORMANCE EVALUATION D.Varying the tag density This section describes the experimental settings and results. In Fig.11 and Fig.12,we respectively show the execution The experimental data in different conditions have been used time and the localization error under different tag densities. to analyse the performances of the above algorithms.We use In Fig.11,we can find that as the tag density increases,each the overall execution time and the accuracy of localisation as protocol's execution time increases.This is because the larger the performance metrics. tag density p leads to the larger number of tags identified by the reader,which increases more execution time.However, A.Experiments Setup when the tag density is large enough(eg.8tags/m,10tags/m), As shown in Fig.8,we deploy the tags on the shelf.The the execution time of CMP and DVP almost remain the same, width and the height of the deployment area equal to 5m and because our proposed protocols use the optimal power to 2m,respectively.The tags are deployed in 5 rows and the tag scan the tags.When the tag density increases,the effective category information is randomly generated.The tags used in interrogation region decrease,while the number of identified the experiment are ALN-9662(HiggsTM3)and the reader is tags almost remains the same.Therefore,the execution time ALR-9900+,which runs on global EPC Gen 2 platform.The of CMP is only 58%of that of CCP.while the execution antenna is Alien ALR Antenna.The antenna faces to the shelf,time of DVP is only 60%of that RP.Based on Fig.12.as the and the distance between the antenna and the shelf is 1m.In tag density increases,the localization error decreases.This is default situation we fix the reader towards the center of the because the variance of the obtained tag information(tag size shelf,and scan the tags for 50 query cycles repetitively. in each category.RSSI of the tags)decreases,when the tag B.Comparision of execution time density decreases. Based on Fig.9,we can conclude that our proposed proto- cols(CMP,DVP)have better performances than the baseline E.Varying the distribution of categories protocols(CCP,RP).It is because that our proposed protocols Fig.13 and Fig.14 show the execution time and the local- scan the tags with the optimal power instead of the default ization error under different category's widths.As shown in 146is an intermediate result in algorithm 1. By setting the power to p0 and scan the tags, we get the identified categories Ci and the RSSI value si,j of each tag in the category Ci. Through the training data set Td, we can get the distance di,j of the tag to the center of the detection region. Take the position xi,s+xi,e 2 of the category’s midpoint as the tag’s position, then the estimation position xˆi calculated by the category Ci is xi,s+xi,e 2 ± ni j=1 di,j ni . The category contains the minimal di,j is the most possible category which contains the real position x0, we can ensure the relative positions of the categories. Then the estimation position xˆ is k i=1 xˆi k . Algorithm 3 Distance Voting based protocol 1: Call Alg.1 to compute the tag density ρ and the optimal value of power p0. 2: Get the identified categories Ci = {xi,s, xi,e, li} and the RSSI value si,j of each tag in the category Ci. 3: Through the training data Td, get the distance di,j of the tag to the center of the detection region. 4: Find the category C0 which has the minimal di,j . 5: for i ∈ {1, 2...k}, j ∈ {1, 2..ni} do 6: if xi,s > x0,e then xˆi = xi,s+xi,e 2 − ni j=1 di,j ni endif 7: if xi,e < x0,s then xˆi = xi,s+xi,e 2 + ni j=1 di,j ni endif 8: end for 9: compute the xˆ as follows: xˆ = k i=1 xˆi k . 10: OUTPUT: the estimation position xˆ. 3) Analysis: This method improves the reliablity of the baseline RSSI weighted average based method. By getting the corresponding distance value of the RSSI, we make the algorithm more robust. VII. PERFORMANCE EVALUATION This section describes the experimental settings and results. The experimental data in different conditions have been used to analyse the performances of the above algorithms. We use the overall execution time and the accuracy of localisation as the performance metrics. A. Experiments Setup As shown in Fig.8, we deploy the tags on the shelf. The width and the height of the deployment area equal to 5m and 2m, respectively. The tags are deployed in 5 rows and the tag category information is randomly generated. The tags used in the experiment are ALN-9662(HiggsTM3 ) and the reader is ALR-9900+, which runs on global EPC Gen 2 platform. The antenna is Alien ALR Antenna. The antenna faces to the shelf, and the distance between the antenna and the shelf is 1m. In default situation we fix the reader towards the center of the shelf, and scan the tags for 50 query cycles repetitively. B. Comparision of execution time Based on Fig. 9, we can conclude that our proposed proto￾cols (CMP, DVP) have better performances than the baseline protocols (CCP, RP). It is because that our proposed protocols scan the tags with the optimal power instead of the default Fig. 8. Experiments setup power used by the baseline protocols. Usually, the larger power means the larger detected tag size, which leads to more execution time. In Fig.9, the tag density is 10tags/m and the average width of categories is 0.5m, the execution time of CMP is only 62% of that of CCP, while the execution time of DVP is only of 57% of that of RP. In regard to the proposed protocols, DVP has larger execution time than CMP, because DVP needs to get the tag IDs and the RSSI value of each tags as well. C. Accuracy of Localization From Fig.10, we can find that all our proposed protocols have higher accuracies than the baseline protocol. Besides, the proposed protocols are more reliable than the baseline protocols with lower localization error. This is because our proposed protocol (CMP, DVP) respectively use the category matching method and the distance voting method to select the k nearest categories to locate the target. They reduce the error caused by the variances of tag distributions. When the tag density is 10tags/m and the average category width is 0.5m, the localization error of CMP is only 75% of that of CCP, while the localization error of DVP is only 73% of that of RP. D. Varying the tag density In Fig.11 and Fig.12, we respectively show the execution time and the localization error under different tag densities. In Fig.11, we can find that as the tag density increases, each protocol’s execution time increases. This is because the larger tag density ρ leads to the larger number of tags identified by the reader, which increases more execution time. However, when the tag density is large enough (eg. 8tags/m, 10tags/m), the execution time of CMP and DVP almost remain the same, because our proposed protocols use the optimal power to scan the tags. When the tag density increases, the effective interrogation region decrease, while the number of identified tags almost remains the same. Therefore, the execution time of CMP is only 58% of that of CCP, while the execution time of DVP is only 60% of that RP. Based on Fig.12, as the tag density increases, the localization error decreases. This is because the variance of the obtained tag information (tag size in each category, RSSI of the tags) decreases, when the tag density decreases. E. Varying the distribution of categories Fig.13 and Fig.14 show the execution time and the local￾ization error under different category’s widths. As shown in IEEE ICC 2014 - Ad-hoc and Sensor Networking Symposium 146