RFID systems [28] are widely deployed to label and track

Transcription

1 IEEE/ACM TRANSACTIONS ON NETWORKING 1 PHY-Tree: Physical Layer Tree-Based RFID Identification Yuxiao Hou and Yuanqing Zheng Abstract Tree-based RFID identification adopts a binary-tree structure to collect IDs of an unknown set. Tag IDs locate at the leaf nodes and the reader queries through intermediate tree nodes and converges to these IDs using feedback from tag responses. Existing works cannot function well under the scenario of nonuniform ID distribution as they ignore those ID distribution information hidden in the physical-layer signal of colliding tags. Different from them, we introduce PHY-Tree, a novel treebased scheme that collects two types of information regarding ID distribution from every encountered colliding signal. First, we can detect if all colliding tags send the same bit content at each bit index by looking into inherent temporal features of the tag modulation schemes. If such resonant states are detected, either left or right branch of a certain sub-tree can be trimmed horizontally. Second, we estimate the number of colliding tags in a slot by computing a related metric defined over the signal s constellation map, based on which nodes in the same layers of a certain sub-tree can be skipped vertically. We thus call the two types of information as horizontal and vertical info. Evaluations from both experiments and simulations demonstrate that PHY-Tree outperforms the state-of-the-art schemes by at least Index Terms RFID, tree-based identification, physical layer. I. INTRODUCTION RFID systems [28] are widely deployed to label and track items in various applications, such as inventory management [30], [32], access control [8], human-machine interaction [27], localization and mobility tracking [24], [26], [29]. One fundamental operation in RFID systems is to read tag IDs (a.k.a., RFID identification). Two major types of RFID identification schemes are ALOHA-based and tree-based. In ALOHA based schemes [18], [31], each tag randomly selects a time slot and responds to reader s query, leading to frequent tag collisions and low communication efficiency. In contrast, tree based schemes allow readers to issue binary prefixes for tags to match their IDs with. Previous works have shown that tree-based schemes provide more stable identification performance but incur more reader-tag interactions [22]. Existing MAC-layer tree-based Manuscript received March 30, 2017; revised August 7, 2017 and November 25, 2017; accepted January 7, 2018; approved by IEEE/ACM TRANSACTIONS ON NETWORKING Editor X.-Y. Li. This work was supported in part by the Hong Kong ECS under Grant PolyU /15E and in part by the Hong Kong PolyU under Grant G-YBMT and Grant G-YBXX. A preliminary version of this work was published in Proceedings of IEEE INFOCOM 2017 [10]. (Corresponding author: Yuxiao Hou.) The authors are with the Department of Computing, The Hong Kong Polytechnic University, Hong Kong ( s120003@e.ntu.edu.sg). Digital Object Identifier /TNET works [13], [14], [16], [20], [22] deliver their optimal performances only when the target tag set renders uniform ID distribution, which rarely occurs in practical scenarios. In order to improve the performance of tree-based identification, we explore how to infer local ID distribution from physical layer signals. When multiple tags collide, we can detect whether all responding tags reply with the same bit at each bit index by combining prior knowledge of tag coding scheme and physical layer patterns. For example, if the query prefix is 0 and the all-0 state is detected at index 3, the reader can infer that no tag has replied with the prefix pattern 0*1 (where * represents any bit value) and skip querying prefixes matching the pattern 0*1. In the binary query tree, this is equivalent to pruning the right branch of the layer-2 nodes rooted at 0. Since this physical layer information tells whether left or right branches of certain tree nodes is empty and thus could be skipped safely, we call such physical layer information horizontal information. In addition, based on collision patterns in the physical layer, we can roughly infer the number of responding tags in each query. For example, if 4 tags respond when the reader queries prefix 01, instead of appending only one bit to prefix 01, the reader can directly append two bits to the prefix and query 4 new prefixes (0100, 0101, 0110, 0111) to directly resolve the collision. In the binary query tree, the reader skips two children nodes (i.e, 010 and 011) and directly jumps down to the 4 grandchildren nodes. We call such physical layer information vertical information. We find that by leveraging both types of physical layer information, it is possible to skip many unnecessary queries during the overall identification process. In this paper, we propose PHY-Tree, a novel tree-based scheme that extracts two types of information on ID distribution (i.e., horizontal and vertical information) from every encountered colliding signal and utilizes them to guide a more efficient query over the binary tree. First, by looking into inherent temporal features of the tag modulation scheme, we can detect whether concurrent colliding tags backscatter with the same bit content at each bit index (horizontal info). Second, we can also estimate the number of colliding tags in a slot without extra communication overhead, by computing a related metric that is defined over the signal s constellation map (vertical info). By accumulating these information along the whole identification process, a great amount of information on ID distribution can be obtained to avoid many unnecessary queries in comparison to existing MAC-only tree-based schemes this is the reason why PHY-Tree outperforms them IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See for more information.

2 2 IEEE/ACM TRANSACTIONS ON NETWORKING Fig. 1. Illustration on how the proposed PHY-Tree scheme accelerate identification performance compared to classic Query Tree scheme. (a) Query Tree takes 20 queryies to identify all 8 tag IDs. (b) PHY-Tree takes only 6 queries to identify all 8 tag IDs. by nature. We also design a mechanism to compensate for errors in the obtained physical layer information and ensure that all tag IDs in the set are correctly identified. We conduct experiments on USRP/WISP testbed and perform extensive trace-driven simulations to evaluate PHY-Tree. The evaluation results show that PHY-Tree outperforms state-of-the-art tree-based identification scheme by II. BACKGROUND &MOTIVATION A. RFID Backscatter Passive tags are of small size and battery-free. They are generally used in UHF RFID systems that operate in the range from 860MHz to 960MHz. RFID reader issues continuous wave (CW) onto RFID tags and tags can transmit data by either reflecting or absorbing CW. In other words, each tag has two transmission states, HIGH (bit 1) and LOW (bit 0) [9], [15]. To increase the robustness against channel degradation and maintain a high decoding rate on the reader side, certain encoding scheme is imposed to the data transmitted by a tag. EPC Class 1 Gen 2 (EPC-C1G2) standard [1] provides several encoding schemes including FM0 and Miller-based coding. When k tags collide in the same time slot, two states of each tag (i.e., HIGH and LOW) combine linearly in the wireless channel. In principle, there are 2 k states in the mixed signal (in complex values). Among 2 k states, two of them are worth noting: all-0 and all-1 state. The two states occur when the k tags transmit bit 0 or bit 1 at the same time. We denote the two states as resonant states, to distinguish them from mixed states. If we plot the received complex signal in 2D plane, we observe many clusters in the constellation map. For each cluster, its centroid represents a combined state of k tags and its radius indicates the power of noise from the wireless channel. Ideally we can observe 2 k clusters (as shown in Figure 2(a) and Figure 2(b)). When k increases, it becomes more challenging to accurately infer k from the number of clusters, as shown Figure 2(c) where 4 tags respond simultaneously. Detailed experiment configurations are described in Section 7.1. B. RFID Identification The task of RFID identification is to collect all IDs of an unknown tag set. ALOHA based schemes [18], [31] require tags to reply randomly in any time slot of a frame and hence suffer from repetitive collisions. On the other hand, in tree based schemes the reader issues a series of binary prefixes and only those tags whose IDs match the issued prefixes would reply. Consequently, collisions are gradually reduced during such reader-tag interactions, and finally all tags can be identified. We explain some basic concepts in tree-based schemes and illustrate the key intuition of our design in Figure 1(a). Each node represents a binary prefix in the tree and is classified into one of the following three types: collision, singleton and empty, depending on the number of responses upon queried. We find that 8 tags are in the bottom layer of the tree (in square shape), whose IDs can be read by traversing from the root to the tags. For instance, the ID of the 3 rd tag from left-most is A tag ID is an unsigned binary sequence that can map to a decimal value. For a given tag set, if we rank tag IDs according to their corresponding decimal values in ascendant or descendant order and regard the difference value between two neighboring tag IDs as a random variable, we could tell if the global ID distribution follows any typical statistical distributions, e.g., normal distribution. Many tree-based schemes [13], [16], [22] are proposed to reduce the number of query prefixes (represented as black nodes in Figure 1). Although their ways of traversing through the tree differ from each other, one common feature in these works is that they append only one bit to the current queried prefix to resolve collisions, if any. C. Motivation Despite the plenty amount of tree-based identification works, they assume either explicitly or implicitly on certain specific tag ID distributions and hence cannot adapt to any arbitrary ID set without a clear distribution. For instance, Smart Trend Traveral [16] performs efficiently when tag IDs are grouped into batches, in each of which continuous ID values present. The current state-of-the-art work Tree Hopping [22] computes an optimal tree layer for each subtree query by assuming tag IDs are uniformly distributed. In contrast, our proposed scheme does not rely on any specific tag ID distribution to operate efficiently. To optimize the query, we estimate the local ID distribution in each query sub-tree by extracting two types of physical layer information, namely horizontal and vertical information. We note that physical layer hints are obtained directly from the physical layer without extra communication overhead. As the hints are accumulated

3 HOU AND ZHENG: PHY-TREE: PHYSICAL LAYER TREE-BASED RFID IDENTIFICATION 3 Fig. 2. Ideally, 2 k number of clusters appear in the constellation map when k tags collide as shown in (a) for k =3and (b) for k =4. However, in practice clusters may overlap and less than 2 k clusters could be observed as shown in (c) for k =4. along the identification process, we could gradually refine the queries and ultimately converge to exact position of each tag ID. We use an illustrative example to explain how horizontal and vertical information could be utilized in tree-based identification. In Figure 1(b) each of 8 tag IDs is labeled with a square in the bottom layer. We start querying with the two nodes in the first layer: 0 and 1. When 0 is queried and the 4 tags in the sub-tree reply, the constellation map may exhibit as shown in Figure 2(b). In this case, the reader can leverage such collision hints, and skip one query node and jump down vertically. As such, two nodes (i.e., 00 and 01) are skipped and marked as V skip (vertical skip). Meanwhile, the reader also obtains the horizontal information that no tags reply with prefix 01 since the 4 responding tags have the same bit 0 at index 2. Consequently, the node 01 and its descendants are skipped for query and marked as H skip (horizontal skip). It is noted that the skipped node 01 is the result of both skip types and marked as H + V skip. Similarly, when 1 is queried, the reader obtains both horizontal and vertical information, which respectively indicate that 4 tags collide and no tags reply with the prefix pattern 1*0 (where * represent a bit value). Thus, the reader further skips the nodes 10 and 11 and two sub-trees rooted from node 100 and 110. Next the reader queries the non-skipped nodes in the third layer (i.e., 000, 001, 101, 111). Upon node 000 is queried, the reader learns that only 2 tags collide and a mixed state appears at bit index 4. For this combination of horizontal and vertical information, the only possibility is that 0000 and 0001 coexist. Thus the reader can directly identify two tag IDs without explicit queries of 0000 and Compared with Figure 1(a) where QT takes 20 queries (black solid circles or squares) to identify all 8 tag IDs, PHY-Tree takes only 6 queries thanks to H skip and V skip. To conclude, PHY-Tree improves the performance over existing MAC layer schemes as it obtains extra physical layer hints on the local ID distribution in each tag response slot and accumulates them along the whole identification process. III. HORIZONTAL INFORMATION A. Horizontal Information In Figure 3, In FM0 coding, a whole bit duration can be divided into the first half and second half. In each half-bit duration, the state of an individual tag i can have two options: Fig. 3. FM0 and Miller-2 coding of a common binary sequence Fig. 4. Colliding signal of two tags with different IDs using FM0 coding. HIGH i and LOW i. Depending on the bit content, a state flipping occurs at the boundary of the two half-bit duration for bit 0 and does not happen for bit 1. We can thus infer that if bit 0 is sent, the latter half of this bit and of its previous bit should have the same state. The above patterns can be extended to the mix signal of multiple tags, as shown in Figure 4. Specifically, no state flipping occurs in the mid of a bit in all-1 state the first and second half of bit 1 have the same state. We see that when both tags send bit 1 (i.e., index 2, 3, and 6), no state flips happens in the corresponding indexes in collision signals. For a bit of all-0 state, the latter half of this bit and of its previous bit are the same. We see that when both tags send bit 0 (e.g., index 5 and 8), the latter states remain the same in the corresponding indexes in collision signals. Mixed states do not exhibit these features and can be distinguished from resonant states. Next we check if resonant states can also be observed in Miller coding. Figure 3 shows an example of Miller-2 coding. We observe that for bit 1, the second and third quarters of the signal have the same state. For bit 0, we observe that the states of the first and third quarter of bit 0 are the same. These two patterns can be easily generalized for Miller-M coding with M =4, 8. In the rest of this paper, we focus on FM0 coding.

4 4 IEEE/ACM TRANSACTIONS ON NETWORKING TABLE I 3 8-BIT TAGS IDS & STATE OF EACH BIT Fig. 5. Constellation map of 3 tags. Their 10-bit IDs include all 8 bit states. (a) Half-bit cluster. (b) Overall cluster. By detecting resonant states from the collision signal, the reader can avoid querying unnecessary prefixes. We note that even if a resonant state is detected at an index not right after the prefix, we still can omit some unnecessary queries. For example, in Figure 1(b), if prefix 1 is queried and an all-1 state is detected at index 3, the reader can skip the nodes matching the prefix pattern 1*0 ; in this case, the nodes 100 and 110 and their descendant nodes can be skipped. B. Robust Detection Algorithm In order to detect the horizon information, we need to compare whether two half-bit states are the same so as to detect both resonant states. To perform the comparison in the complex tag response signal, we can find the corresponding complex values for the two states and compare their real and imaginary parts. However, the backscatter signals may suffer from noises which makes obtaining accurate horizon information more challenging. Next we design a robust resonant state detection algorithm. In physical layer constellation map, since each cluster represents a combination state from all responding tags, we can infer if states in two half bits are the same by judging whether their corresponding clusters are overlapping with each other. Figure 5(a) shows an example of overlapping clusters when 3 tags reply simultaneously. Each labeled cluster corresponds to samples in a half bit. The 3 tags transmit the following 10-bit sequences: , and , which include all 8 possible states, i.e., from 000 to 111. At the bottom-right corner of Figure 5(a), the latter half of state 000 at index 3 (in brown circle) shares a large overlapping region as the latter half of state 111 at index 2 (in green solid square). In fact, we can distinguish between the first and second half of a bit from the tag response signal in the time domain. Moreover, both the frontier and latter half-bit clusters of state 111 at index 1 (in green solid square) and index 10 (in black dot) are almost overlapping as well. In contrast, states other than 000 or 111 do not have similar overlapping patterns. Figure 5(b) shows the overall constellation map. To quantify the notion of cluster overlapping, we first compute the centroid of two clusters as the average of corresponding samples. Two clusters are judged as overlapping if the distance between their centroids is below a threshold: (x2 x 1 ) 2 +(y 2 y 1 ) 2 c r 0, (1) where (x 1,y 1 ), (x 2,y 2 ) are the coordinates of two centroids and r 0 is the cluster radius. c is tunable and empirically set to 2.5 in Section 7. We can approximate r 0 as the L2-norm of all samples in a cluster: r 0 = q (x i x) 2 +(y i ȳ) 2. (2) i=1 In real-time operation, the reader can obtain r 0 by collecting background noise. C. Extension to Mixed States We wonder if similar horizontal information on mixed states could be extracted as well. Suppose 3 tags respond to the issued querying prefix 01. Table 1 lists the 8-bit IDs of three tags. In the row of STATE, A 0 and A 1 represent two resonant states: all-zero and all-one. M 1, M 2 and M 3 are three different mixed states, corresponding to 110, 001, and 101 respectively. We note that state M 1 occurs at both index 2 and 3. If this information is correctly detected, after the reader encounters such a 3-tag collision with prefix 01, among its four children prefixes ( 0100, 0101, 0110 and 0111 ), the reader can safely skip 0101 and 0110 for subsequent queries, as each of the 3 matching tags has the same bit at index 2 and 3 (i.e., 0100 or 0111 ); otherwise the states at index 2 and 3 should differ. Since in FM0 coding scheme state flipping always occurs at the bit boundaries, the bit information lies in whether an extra state flipping occurs in the middle of the whole bit duration. One step further, by subtracting the state of the latter half-bit at bit index j 1 from that at bit index j (denoted as the state difference at index j), we can infer its bit content. It is noted that each half-bit duration can take one of the two states: HIGH i or LOW i. Hence, for bit 1, the state difference can be either HIGH i LOW i or LOW i HIGH i, depending on the context as FM0 coding has memory; for bit 0, the state difference is 0 - this is equivalent to the judgement rule of the two same halt-bit states for bit 0 in Section 3.1. To ease the illustration of this concept, we adopt the same half-bit cluster representation in the constellation map as in Section 3.2. We denote the centroid of the latter half-bit cluster of a certain tag i at bit index j as c ij. We use the following state transition vector v ij to characterize the state difference of tag i at index j (j >0): v ij = c ij c i,j 1 (3) If the bit content b ij is 0, then v ij = 0; ifb ij =1,then v ij = v i0 or v ij = v i0,wherev i0 = c HIGH c LOW.

5 HOU AND ZHENG: PHY-TREE: PHYSICAL LAYER TREE-BASED RFID IDENTIFICATION 5 Fig. 6. State transition vectors for 2-tag and 3-tag collisions. Each mixed state corresponds to at least one pair of transition vectors with same amplitude and opposite directions. (a) When 2 tags collide, k 1 =1for state 01 and 10. Each mixed state maps to 2 vectors (1 pair). (b) When 3 tags collide, k 1 =2 for state 110, which maps to 2 pairs of symmetric vectors. We use c HIGH and c LOW to represent the two centroids of the half-bit clusters that correspond to state HIGH i and state LOW i. We also use the footnote 0, instead of j, to indicate that the two states of tag i do not vary over time. A mixed state at bit index j for multiple colliding tags indicates at least one tag i 1 transmits bit 0 and meanwhile at least another tag i 2 transmits bit 1. To ease the discussion we focus first on the case where only two tags collide. We denote the corresponding synthesized state transition vector for the colliding signal at index j as v sj = v 1j + v 2j. If tag 1 transmits bit 0 and tag 2 transmits bit 1 at both index j 1 and index j 2 (j 1,j 2 > 0), then we can infer that v s,j1 and v s,j2 take either v 20 or v 20. Similarly if tag 1 transmits bit 1 and tag 2 transmits bit 0, then v s,j1 and v s,j2 take either v 10 or v 10. An equivalent graph illustration is shown in Figure 6(a), where 01 and 10 represent the two mixed states. Specifically, 01 indicates that tag 1 transmits 0 and tag 2 transmits 1 and vice versa for 10. As we can observe from Figure 6(a), the state transition vector of each mixed state can take two values, which have the same amplitude and completely opposite directions. As such, we can judge if the mixed states at index j 1 and j 2 (j 1,j 2 > 0) arethesame by comparing if the corresponding state transition vectors (i.e., v s,j1 and v s,j2 ) have the same amplitude, and meanwhile either same or opposite directions (or angles). The equivalent mathematical expressions are: v s,j1 = v s,j2, v s,j1 = v s,j2 or v s,j1 = v s,j2. (4) To tolerate noise in practice, we use similar threshold-based judgement rules as in Section 3.2, but in vector expression instead: v s,j1 v s,j2 z r, v s,j1 v s,j2 z a, (5) where z r and z a are thresholds that can be trained off-line. As in Eq.1, z r can also take the form of c r 0. v s,j1, v s,j2,andz a can take radian values in the range [0, 2π). When k tags collide at index j, suppose the first k 1 of them send bit 1 and the rest send bit 0. Hence, v sj = k i=1 v ij = k 1 i=1 v ij + k i=k v 1+1 ij. As discussed above, v ij = 0 when b ij = 0. Hence, v sj can be expressed as k1 i=1 v ij = k 1 i=1 ±v i0. Figure 6(b) plots the case where k =3and k 1 =2.Since v sj = ±v 10 ± v 20. The mixed state 110 corresponds to 4 possible transition vector values, which are marked as dashed light blue arrows in Figure 6(b). The four v sj values can be divided into two groups, in each of which a pair of vectors with same amplitude and opposite angles exist. Hence, we can safely generalize that Eq.5 can still be used to detect if two mixed states are the same, when k>2. The above horizontal information on mixed states can be integrated together with the one on resonant states in the ultimate PHY-Tree algorithm, by simply adding the corresponding rule to skip tree nodes horizontally after the inference vector (defined in Section 5.1) is extracted from each colliding signal. Although the condition for detecting the same mixed states in Eq. 5 bases on the theoretical assumption on linear combination of individual backscatter signals, PHY-Tree is also robust to errors caused by those non-linear factors (e.g., environmental multipath effect, nearby dynamic objects). Specifically, while false positives can be efficiently compensated through the error compensation scheme (stated in Section 5.2), false negatives render the loss in chance of skipping a few intermediate tree nodes, at most. Generally, such an extra type of horizontal information can facilitate the whole tree-based identification process. We leave the performance study and further exploration of this horizontal information for future works. IV. VERTICAL INFORMATION The number of responding tags k in a slot can help the reader to adjust the prefixes in the following queries. For example, if 4 tags respond to the prefix 0, 2 bits should be appended for the next query. In ALOHA schemes [31]: the optimum slot efficiency is achieved when the number of slots equals to the number of tags. Similarly, to accommodate k tags in the sub-tree, we need to append log 2 k bits to the previous query. Many cardinality estimation techniques [4], [19], [21] can estimate k with high accuracy but they involve extra query overhead. Other works [2], [9], [12] estimate k by counting the number of visible clusters in the constellation map of the tag response signal. Although these works are lightweight, they are not scalable to large k values. This is because when k becomes large, individual clusters have the tendency to overlap with each other and consequently, cluster counting based method no longer performs well. As a result, it is hard to accurately count the exact number of responding tags when the number of tags increases. Instead of aiming at the exact k, we find that it suffices for PHY-Tree to estimate the logarithmic scale of k to make an intelligent bit appending decision in subsequent queries. To this end, we explore a new metric defined over the signalâł s constellation map to estimate the scale of k. A. Intuition & Definition Since an indicator of k only needs to monotonically increase or decrease over k, the area occupied by the physical

6 6 IEEE/ACM TRANSACTIONS ON NETWORKING Fig. 7. Experiment results. (a) Expectation. (b) Standard deviation. layer symbols is a good candidate because: 1) it might increase when k increases; 2) it is robust to overlapping clusters for large k. Based on the above observation, we define an indicator of k, named effective area (EA), as follows. First we find the smallest rectangular region that contains all data samples in the constellation map and divide it into small square grids. Next we count the number of samples in each grid and set a threshold to differentiate noise grids from signal grids the grids with the number of samples below the threshold are regarded as noise. EA is computed as the multiplication of the number of signal grids and the area of the grid unit and can be explained as the summation of areas of all signal grids. B. Basic Observations From Experiment Results We conduct experiments on USRP/WISP testbed (shown in Figure 9) to validate whether EA is a good indicator of k. We configure multiple WISP tags to reply concurrently with random number (RN) sequences to single reader queries. We collect more than 500 concurrent k-tag response signals. As it is hard for us to obtain traces with k > 4 from our testbed, we could easily generate them by adding up available experimental traces. For instance, we can synthesize a trace of k =6by adding up two traces of k =2and of k =4,which also increases the diversity of synthesized traces. Figure 7(a) plots EA versus k based on the collected and synthesized traces. To compute EA, we discover that when the grid size is comparable to the average cluster radius, which is as detailed in Section 4.3 (In the real implementation it is set as ). We also configure an adaptive grid density threshold to be 13% of the length of each trace. After then, we compute EA values for corresponding traces of the same k and plot the average EA in Figure 7(a), where we observe that EA increases exponentially over k. We also plot the standard deviation of EA versus k in Figure 7(b). We find though the standard deviation of EA also increases when k increases, it is small compared to the corresponding mean value of EA, meaning that EA is a stable indicator of k across diverse traces. C. Extension to Larger-Scale Simulation Since we aim to validate the pattern of EA over k in larger scale and it is non-trivial to obtain high-quality experimental traces with k>4, we turn to simulate k-tag collision signals, from which we compute the EA metric. One important concern for using simulated traces is that we wonder if the pattern of EA over k does not rely on or hardly gets affected by any unpredictable environmental factors, but a regular and inherent phenomenon with respect to the signals themselves. To better emulate the experimental traces, we simulate the baseband response signal of a random tag ID according to the statistics of SNR profiles of those experimental traces. First, we state how to compute the SNR profile of an available experimental trace, which contains k-tag collision signal in one query slot. We refer to Figure 2(a) for constellation map of a 3-tag collision signal, which visually contains 8 clusters. We perform classic DBSCAN clustering algorithm [6] on each experimental trace and compute the radius and centroid of each cluster. Intuitively the radius of a cluster characterizes the expansion of a signal state and naturally features the square root of the noise power N 0. Hence, we compute the average radius of all clusters as N 0. Next, the signal amplitude of one of the colliding tags is characterized by the distance between centroids of corresponding cluster pairs (depending on the global states of all colliding tags). In practice, the colliding signal suffers from energy degradation and phase drifts, due to factors like nonideal transmission channels. As such, we choose a reference cluster from all, which renders the shortest Euclidean distance to the origin, and calibrate the new origin of the constellation map as its centroid. After then, the average signal amplitude (or the square root of signal power S) is computed as: 1 S = z i z 0 (6) C i where C is the number of clusters, z 0 and z i are centroid coordinations of the reference cluster and of cluster i. Itis noted that the reference cluster is self-inclusive when computing S using Eq.6; in other words, z i can be z 0. Following above methods, we display the computation results as follows. The average N 0 over all 557 collected traces is One issue to be noted when computing the signal amplitude is that k individual response signals and their signal powers add onto each other for k>1. Hence, the average amplitude over traces of the same k need to be normalized by k first to obtain the average individual amplitude over each k. The ultimate S is computed as the weighted sum of those normalized amplitudes over the number of traces for each k. We note that the signal power varies a lot from traces to traces, due to kinds of environmental factors. Consequently, we also compute the standard deviation of S (similar normalization and weighted summation are included as well), so that we can emulate the signal power diversity when generating traces. After all, the mean and standard deviation of S is and , respectively. We can immediately infer the average SNR of experimental traces is about 19.53dB. We use the above statistics to generate k-tag collision signals, compute their EA values for each k (using the same grid-related settings as for experimental traces), and plot the pattern of EA over k in Figure 8, from which we observe a similar tendency as in Figure 7. This validates the robustness of EA as a k-indicator even when k is large.

7 HOU AND ZHENG: PHY-TREE: PHYSICAL LAYER TREE-BASED RFID IDENTIFICATION 7 TABLE II MAPPING TABLE BETWEEN EA AND k Fig. 8. Simulation results. (a) Expectation. (b) Standard deviation. D. Appending Rule The number of observed clusters in the constellation map is always upper bounded by the number of data samples N s in a response slot. Each of our experimental traces contains 6000 to 8000 samples, which allows to estimate k up to 12. In fact, it makes no big difference to infer 8 or more than 8 (but less than 16) colliding replies if k is used for tree-based identification. Specifically, we need to append log 2 k bits to the current prefix for further queries, if the length of the prefix after appending does not exceed the tag ID length. Following this rule, 4 bits will be appended if k is detected to be in the range [9, 16]. Hence, if we can infer 8 tags at most and the corresponding EA is denoted as a 0, we would append 4 bits for any measured a satisfying a>a 0. One may argue that more sophisticated rules can be applied to determine the number of appended bits. For k = 9, appending 3 bits may resolve most collisions but appending 4 bits incur some inefficient queries to which no tag replies. Though appending 3 bits seems more optimal than appending 4 bits, it does not fundamentally improve the identification performance. First, our mapping model between EA and k is a coarse one and originated from experimental observations. Second, in general tree-based schemes, a rough estimation of k for each reader query without incurring extra overhead is sufficient to bring promising performance gain for the whole query process. It is noted that the seemingly small scalability in estimating k does not indicate its limited gain in the whole identification process. We can estimate k for each tree node and skip several tree nodes if k>1. Such individual gain will accumulate in many query rounds of the identification process and become significant ultimately. E. Model EA Over k We propose a non-parametric model for the mapping between EA and k. Since inferring k up to 8 suffices (according to Section 4.4), we adopt a range-based mapping model for each k [1, 8]. If the measured EA falls in a certain range, the associated k can be inferred. Such a model is robust because the range estimation error for any k 1 is seldom propagated to the estimation for another k 2. To train the range-based model 8 tags are chosen from the tag set and multiple traces are collected for each k. After that, the expectation E and standard deviation σ of EA over these traces can be collected as shown in Figure 8. The minimum/maximum values of EA over k can be obtained using E ± wσ, wherew is a tunable parameter and set as 2 empirically. One issue here is that ranges of neighboring ks may gap or overlap. Thus, it is important to smooth the range boundaries between the upper bound of k and the low bound of k +1, by simply using the same average value. Table 2 shows the boundary-smoothed EA range for k [1, 8] with respect to our simulation traces. This table can serve as the real-time detection of the vertical information, since the overheads involved in both EA computing and the table searching is at most linear to the sample length in a slot. For instance, if EA falls into [ , ), k is estimated to be 6 according to the mapping between EA and k in Table 2. It is noted that we can train such table off-line in practical scenarios, where a targeted tag set may contain hundreds or thousands of tags. It is a one-time effort and does not impact the performance of subsequent identification process. V. PHY-Tree PROTOCOL We propose an efficient RFID identification scheme by exploiting both horizontal and vertical information. A. A Basic Protocol Suppose a target set contains M tags, each has an ID with the length of L bits. We denote the ID of the tag j as ID j = b L 1 b L 2...b 1 b 0,whereb L 1 is MSB and b 0 is LSB. When the reader queries with an l-bit prefix Q = q l 1 q l 2...q 1 q 0 (1 l L), tag j checks if the first l bits of its ID, b L 1 b L 2...b L l, match Q. If they match, the tag j replies to the reader with its ID. The behaviors of reader and tag in the basic PHY-Tree scheme are described in Algorithm 1 and Algorithm 2, respectively. The reader starts the query from the first layer of the binary tree (i.e., Q = 0 or Q = 1). Upon receiving the colliding physical layer signal from k tags for the prefix Q, the reader can compute an estimation of the number of replying tags ˆk (vertical information) and obtain an inference vector G = g L 1 g L 2...g 1 g 0 (horizontal information), where g i (i = 0, 1,...,L 1) indicates the bit state at index i. Specifically, g i is set to 0 or 1 if an all-0 or all-1 state is detected respectively and 1 otherwise. Since only tags matching Q respond, we know that g L 1 g L 2...g L l = q l 1 q l 2...q 1 q 0. If collision happens,

8 8 IEEE/ACM TRANSACTIONS ON NETWORKING Algorithm 1 Basic PHY-Tree Algorithm for RFID Reader 1: N { 0, 1 }, N is always automatically sorted in increasing order. 2: while N do 3: Pop out the smallest prefix Q from N. 4: Send the query command with Q to tags. 5: Receive tag responses. 6: 7: Compute G, ˆk, d. h max( log 2ˆk,d+1). 8: if ˆk >1 then 9: Obtain the appended prefix query set Q (of size 2 h ) by appending h bits to Q. Filter Q with G. 10: Add Q into N. 11: else if ˆk =1then 12: Identify a tag ID. 13: else 14: Continue. 15: end if 16: end while Algorithm 2 Basic PHY-Tree Algorithm for RFID Tag 1: while TRUE do 2: Wait for the reader s query command containing query prefix Q = q l 1 q l 2...q 1 q 0. 3: Compare its first l-bit ID (b L 1 b L 2...b L l ) with Q. 4: if b L 1 b L 2...b L l = q l 1 q l 2...q 1 q 0 then 5: Reply to the command with its full ID (b L 1 b L 2...b 1 b 0 ) 6: else 7: Keep silent. 8: end if 9: end while the reader resolves the collision as follows. First it uses ˆk and the appending rules to derive the optimal number of appended bits h. The reader appends h bits to the current queried Q to obtain the appended prefix query set Q for subsequent queries. For example, if h is 3, Q includes prefixes from q l 1 q l 2...q 1 q to q l 1 q l 2...q 1 q Second the reader uses G to filter out unnecessary prefixes from Q. If g L l 3 = 1 in the above example, Q is reduced to only 4 prefixes: q l 1 q l 2...q 1 q 0 001, q l 1 q l 2...q 1 q 0 011, q l 1 q l 2...q 1 q and q l 1 q l 2...q 1 q Generally h is solely determined as log 2ˆk. However, if subsequent d bits are detected as resonant states right after the prefix Q,i.e.,g L l 1...g L l d takes valid values (0 or 1), h is adjusted to be max( log 2ˆk,d+1). We describe the whole query process as follows. In the binary tree representation, the reader queries in the breadthfirst mode. Denote N as the queue of subsequent prefixes for query. N is initialized with 2 prefixes 0 and 1 and is automatically sorted in increasing prefix length. In the case that two prefixes in N have the same length, the one with smaller value is queried first. For example, prefix 010 is queried before prefix 011. The reader first obtains the minimum prefix length l min in N and queries all prefixes with length l min in N. Each time the reader pops out the first prefix in N for query and waits for tag responses. If less than 3 tags reply, the reader continues to pop out a new Q in N for query. Specifically if one or two tags reply to Q, they can be uniquely identified. Otherwise, the reader extracts G, ˆk and d from the colliding signal to compute h. The reader appends h bits to Q to get the initial Q, filters out some prefixes in Q using G and adds the updated Q to N. The reader repeats the above query procedure until N becomes empty. B. Error Compensation The bit state detection algorithm in Section 3 may mistake mixed states for resonant states and guide the reader to ignore some sub-tree nodes of the current prefix. Therefore, some tags in the set are missed by the reader. We need to eliminate this problem to meet the basic requirement of RFID identification. A straightforward solution is to record the list of unqueried prefixes due to the guidance of G for each query in the basic algorithm because the missed tags may reply to these prefixes. The reader may adopts the original QT scheme [13] to query prefixes in this list and identifies the remaining tags, without using either horizontal or vertical information. Specifically, it uses depth-first query method to traverse the whole binary tree and always appends one bit for the next queries if collision is encountered. Although it ensures no tag is missed, this solution counters against the benefit brought by horizontal information and degrades the performance gain. Here we propose a more efficient compensation scheme that could reserve the benefit of horizontal information. The simulation results in Section 7.1 show that the accuracy of detecting resonant states is high under tight synchronization among replying tags. This indicates that a majority of tags in the target set are identified using the basic algorithm. This motivates us to design the following error compensation scheme. After the basic algorithm stops, we mute those tags which have been already identified so that they will not interfere with the second round, where the reader performs a fresh identification using QT, starting from the first layer. Since only a few unidentified tags are left in the second round, it does not take many reader queries to finish the second round, indicating its efficiency. C. Further Improvement in Efficiency By combining horizontal and vertical information in another perspective, we can identify 2 colliding tags without further queries and hence gain extra identification ability. Basically if we can estimate ˆk =2for one query with high accuracy, we can uniquely identify each of two tags with its partial ID, rather than decode the whole ID. We explain the reason as follows. Since each tag owns a unique ID, IDs of the two tags must take different bit values in at least one bit index (e.g., i). In other words, a mixed bit state exists at index i. Our bit state detection algorithm can find out bits with mixed states and its accuracy is quite high when only two tags reply (reported in Section 7.1). The partial IDs for two tags are thus set as the queried prefix concatenated with bit 0 and 1 at index i. We incorporate this feature in the tree-based algorithm by slightly modifying the appending rule used in the basic

9 HOU AND ZHENG: PHY-TREE: PHYSICAL LAYER TREE-BASED RFID IDENTIFICATION 9 protocol in Section 5.1: h =max( log 2ˆk 1,d+1). (7) The term log 2ˆk 1 in Eq.7 reflects the change brought by extra identification ability of 2 tags. It is noted that we cannot extend the partial ID identification to the case in which more than 2 tags reply, due to two reasons. First, the estimation error in the EA-k model increases when k increases, leading to the inaccurate estimation results for k>2. Second, even if k is correctly estimated, we cannot identify these tags definitely as more than k possible state combinations exist in several bits. Fig. 9. Our USRP-N210 and WISP 4.1 testbed. VI. DISCUSSION ON SYNCHRONIZATION When we illustrate two types of physical layer information in Section 3 and 4, we assume that the responding tags in a slot are tightly synchronized. In real world scenarios, tags have large diversities in their response delays due to multiple factors such as manufacturers, types, antenna orientations, etc. In this section, we briefly discuss how imperfect synchronization affects the two types of physical layer information and suggest how to mitigate the impact. On one hand, the imperfect synchronization makes negligible impact on the vertical information. We estimate the number of concurrent replies by processing samples in the entire time slot, rather than samples in individual bit durations. Thus the vertical information does not rely on the internal timings between contiguous bit durations in the response signal. On the other hand, imperfect synchronization degrades the quality of the horizontal information. Specifically, discovering all-0 and all-1 states in a bit requires bit-level synchronization. Without perfect synchronization, we may accidentally detect a resonant state as a mixed state (false negative), which leads to missed opportunities for saving some reader queries, or detect a mixed state as a resonant state (false positive), which results in some tag IDs being not identified. Unlike Buzz [25], which can calibrate for each tag s clock to adjust its starting time offset so that responding tags achieve tight synchronization, we aim to identify all IDs of a completely unknown tag set. Thus, we should not devote any effort to calibration; otherwise such an effort can afford us to collect all tag IDs one by one. We suggest several approaches to mitigate the impact of imperfect synchronization on our scheme. Each tag can use lower frequency for backscattering its ID to reduce the unsynchronization rate, which is the ratio of the maximum starting time offset over the individual bit duration. For example, if 4 symbols are used to encode one bit, the unsynchronization rate can be reduced by half compared to using FM0 coding. Better circuits are expected to be designed for RFID tags to further reduce their unsynchronization rates [5], [7], [17]. Besides, to completely eliminate false positives for the horizontal information, we design an error compensation mechanism in Section 5.2. Additionally, powerful collision recovery schemes like BiGroup [15] (which can decode 4 to 5 collision tags without assuming tight synchronization) can augment our scheme to further improve the overall identification performance. Fig. 10. Optimal range for c in resonant state detection with different ks. VII. EVALUATION A. Microbenchmark In this section, we evaluate the effectiveness of horizontal and vertical information. We collect practical traces from our USRP/WISP testbed in the office environment, as shown in Figure 9. The USRP motherboard is equipped with an RFX900 daughterboard operating at 900Hz UHF band to transmit reader commands and receive backscatter signals from tags. Upon receiving the tag response signal, the USRP reader samples the down-converted baseband signal at a rate of 4 million samples per second. As a WISP tag adopts Miller-4 encoding scheme and its backscatter link frequency is 64kbps, each bit lasts for 0.125ms in the air and occupies 500 baseband samples. These samples can be represented by complex values that contain both quadrature and in-phase components. They are sent to the connected laptop for off-line processing. We configure WISP tags to reply with their random number sequences concurrently to single reader queries. We collect 557 traces in various operation environments. 1) Horizontal Information: We first study the performance of the resonant state detection in obtaining horizontal information. Three important performance metrics are overall accuracy, false positive rate (FPR) and false negative rate (FNR). The overall accuracy refers to the ratio of the number of correct detection of both resonant and mixed states over the total number of bit states. False positive rate refers to the ratio of the number of detecting mixed state as resonant state over the number of actual mixed states. False negative rate refers to the ratio of the number of detecting resonant state as mixed state over the number of actual resonant states. Unless otherwise noted, we repeat an algorithm for 100 runs to obtain its average performance for each setting. We find the optimal c, an important tunable parameter in Eq.1. We vary k from 2 to 8 and design a set of 10-bit tag IDs (which is enough to accommodate 8 different IDs) for each k. With random SNR and fixed noise power of each generated trace, we compute an average overall accuracy for each k. We plot the overall accuracy over c in the range [0.1, 10] in granularity of 0.1 in Figure 10. From this figure, we discover

10 10 IEEE/ACM TRANSACTIONS ON NETWORKING Fig. 11. Performance of the resonant state detection algorithm in obtaining horizontal information when k varies in a wide range and c is fixed to be optimal. (a) Overall accuracy. (b) False positive rate. (c) False negative rate. that for each k, the overall accuracy first increases sharply with c, maintains high in a certain range and decreases slowly with c in the end. We highlight the 0.95-accuracy line in this figure and find an optimal range [1.39, 4.07] of c for all ks. For k =2, the overall accuracy is always 100% for any c, demonstrating the credibility of the partial ID identification of 2 colliding tags. After obtaining the optimal c, we study the accuracy of the resonant state detection with different k. We set tag ID length to 100 to better accommodate more unique tag IDs and each of the all-0, all-1 and mixed state occupies 33.3% of the total 100 bits. We choose 6 values for c from the optimal range (1.5, 2, 2.5, 3, 3.5 and 4). A slight difference from the setting in Figure 10 is that we generate random set of tag IDs in each run given a value of k. We plot the overall accuracy, FPR and FNR over k from 3 to 100 in Figure 11. In Figure 11(a), the trend is that the overall detection accuracy of both resonant and mixed states decreases when k increases. This is natural as collisions from more tags impose more challenges on the detection. Nevertheless, the algorithm achieves around 70% accuracy when 20 tags collide for c = 1.5. From Figure 11(b), we find FPR increases when c increases. We emphasize in Section 5.2 that high FPR of horizontal information is not desired and hence should be as low as possible. As such, we can trade off between the overall accuracy and FPR; if we set c to 2.0, we can obtain 1% FPR and achieves 80% overall accuracy when k is 20. In Section 7.2 when we compare PHY-Tree with existing tree-based identification schemes, we fix c to2inphy-tree. In Figure 11(c), we can see that FNR increases over k and decreases over c, which matches the trend in Figure 11(a). Any non-zero false positive rate would trigger our error compensation scheme. If the number of missing tags after the basic query round is large, the follow-up error compensation round incurs large query overhead and renders our proposed PHY-Tree inefficient. We thus need to study about such overhead to better evaluate our scheme. Two metrics are considered: the percentage of missed tags over the tag set population M in the basic query round and the percentage of queries in the error compensation round over the total number of queries. We evaluate the PHY-Tree algorithm on M tags with randomly generated unique 15-bit IDs, where M varies from 2 6 to Figure 12 plots the result. In Figure 12(a), we find that the average percentage of missed tags is 2.2% for all M values. In Figure 12(b), Fig. 12. Impact of false positive of horizontal information on PHY-Tree. (a) Percentage of missing tags in basic query round. (b) Percentage of NoQ in the error compensation round. we find the percentage of queries in the compensation round is 4% when M>8. Both figures prove the small overhead and high efficiency of our error compensation scheme. 2) Vertical Information: Next we study the accuracy of obtained vertical information. As discussed before, we only need to infer up to 8 tag replies with reasonable accuracy and for k>8, we only need to judge if ˆk >8. We adopt the non parametric model and train the range for each k [1, 8]. The training set contains 100 randomly generated traces and the testing set contains another 100 newly generated traces, for each k. Each tag replies with 100-bit RN sequence. In both the training and testing sets, we create 9-tag collision traces for k>8, which is the most challenging scenario for our EA-k model since EA increases over k. We compare the estimation ˆk with the ground truth k and compute the ratio of the number of correctly inferred traces over the overall number of traces as the accuracy. Figure 13(a) plots the result. It is observed from Figure 13(a) that our EA model achieves 80% accuracy on average. The corresponding accuracy is somehow low for certain k value like 2, 6, 8. This is because of overlapping ranges incurred in the training test: the upper bound of the range for k is larger than the lower bound of the range for k +1. To alleviate this issue, we can set the middle of the overlapping range as the common value for the upper bound of k and lower bound of k +1. In Figure 13(b), we compare the accuracy between EA-k model and SSDA in [9], which can estimate up to 4 colliding tags. We implement SSDA using original settings in [9] and set k in range [1, 4] for fair comparison. We observe from Figure 13(b) that SSDA achieves higher accuracy ( 90%) than our EA-k model does, but the EA-based method is more

11 HOU AND ZHENG: PHY-TREE: PHYSICAL LAYER TREE-BASED RFID IDENTIFICATION 11 Fig. 13. Accuracy comparison between EA-k model and SSDA [9]. (a) Inference accuracy of the EA model for wide-range k. (b) Comparison with SSDA for small-range k. Fig. 14. Comparison of EA-k model and SSDA in time. (a) Trace operation time. (b) CDF of operation time. scalable than SSDA. Actually we can combine both methods for better accuracy. Specifically, we can use both methods to judge whether both estimation results for k are below 4. If it is so, we apply SSDA for better accuracy; otherwise we apply EA model for better scalability. From Figure 14(a), we can infer that EA-based and SSDA methods take around similar computational time over each trace. This can be explained from their similar steps of dividing the plane into square grids, counting the number of samples in each grid and judging if each individual grid contains more than a threshold of samples. According to Figure 14(b), EA-based and SSDA methods take less than 1.8ms to process 90% of all traces, which is due to their linear computational time over the number of samples. In conclusion, EA-based method is as time-efficient as SSDA. B. Comparison With Existing Identification Schemes In this subsection, we compare PHY-Tree with three existing tree-based identification works: QT [13], STT [16] and TH [22]. We consider three typical ID distributions: uniform, normal, and block distribution. While uniform and normal distributions are well-known statistical distributions and adopted in previous works [16], [22], the block distribution often occurs in practical inventory management scenarios. Specifically, a tag set following block ID distribution with block size b is grouped into many subsets, each of which contains b contiguous tag IDs (in decimal values). The gap between neighboring subsets, which is the difference between the smallest ID in the next subset and the largest ID in the previous subset, may be random and differ from each other. Similar to [22], we adopt the average number of queries (NoQ) as the performance metric, which is defined Fig. 15. Detailed study on block ID distribution. (a) Small block size. (b) Large block size. as the ratio of total number of reader queries over tag set population M. Each tag ID is 15-bit long and M varies from 2to2 12. Under uniform distribution, tag IDs are randomly created in the range [1, ]; the probability density function for generating normal ID distribution has a mean of 2 14 and a standard deviation of 2 14 /3; the block size b in block ID distribution is initially set to 10 this requires the number of tags be no less than b. It is noted that each generated tag ID is unique among the whole set and the generated integer is converted to the form of 15-bit-long binary sequence, which is actually used in tree-based identification process. For each M, we repeat the identification process for 100 times and compute the average NoQ as the ultimate performance metric. Figure 16 shows the comparison results under three types of ID distributions. Our observations in Figure 16 are as follows. Under uniform ID distribution, the average number of queries for PHY-Tree is around 1 and thus reduces the number of queries of stateof-the-art TH by The comparison under normal ID distribution is similar to that under uniform one. It is noted that under block ID distribution, STT performs better than stateof-the-art TH scheme, due to its perfect knowledge on the tag ID continuity. Even thoufgh, our proposed PHY-Tree performs much better than STT. In a word, PHY-Tree performs best under all three types of ID distributions, thanks to its fullyutilized PHY info on local ID distributions, which is extracted from every query response signal and could be accumulated along the whole identification process. It is noted that this gain of 1.79 over TH is relative; when the magnitude of the number of tags in a set is large, e.g., in billion scale, the reduction in the absolute number of the necessary query rounds becomes significant. It is also noted that the TH scheme in comparison needs perfect cardinality knowledge as input. In practice, a lightweight cardinality estimation process not only takes more query rounds and increases identification overhead, but also influences the hopping performance. In contrast, the proposed PHY-Tree scheme does not suffer from inaccurate input results of lightweight cardinality estimation and has even larger gains in this sense. We note that the time overhead for computing horizontal and vertical information can be paralleled in real time with the identification process, if this time is no longer than the duration of one tag response slot. Specifically, this time involves computing vertical and horizontal information. On one hand, from Figure 14(b) we discover that 90% of the time for computing

12 12 IEEE/ACM TRANSACTIONS ON NETWORKING Fig. 16. Performance comparison between several schemes under three typical ID distributions. (a) Uniform distribution. (b) Normal distribution. (c) Block distribution with b =10. the vertical information, which is proportional to the number of samples, over a 2ms trace takes less than 1.8ms or so. On the other hand, the time for computing the horizontal info is far less than that for computing the vertical one, as it is proportional to the number of bits in a tag ID. Hence, in the following discussion we mainly focus on the computational time for vertical info. From Figure 14(a), we find that sometimes the computational time can exceed 2ms, i.e., a response slot duration. As it is upper bounded by 4ms, computing the vertical info occupies at most 2 response slots. Under such cases, we can still parallel it, due to the following observations. First, only when the reader reaches to query the subtree nodes of the current node whose vertical info is being processed, the obtained vertical info can be utilized. Second, PHY-Tree adopts a width-first tree query approach, under which the reader would query tree nodes at the same layer first before moving on querying the next lower layer. Hence, it takes at least 2 query rounds for the reader to query the subtree nodes of the current node, even if it locates at the first layer (i.e., 0 and 1 ). Third, since tree nodes at the same layer can be queried in any order without extra overheads, the reader can also adjust the query order in the same layer to accommodate some long processing cases and maximally ensure parallel processing. This also justify the usage of the average NoQ as the performance metric, instead of the usage of the absolute running time. We are particularly interested in how these schemes function in the practical block ID distribution with varying b, which characterizes the level of tag ID continuity. We fix M to 5000 and vary b in the range [10, 100] in granularity of 10 and [100, 1000] in granularity of 100. For each b we create 100 random sets of tag IDs and execute identification schemes on these sets. We plot the average NoQ over b in Figure 15. In Figure 15 we find that STT outperforms TH under block ID distribution. When b is large (e.g., 100), the performance of STT nearly approaches that of PHY-Tree. This is because STT fits best to distributions of long ID continuity. However when b is small, PHY-Tree is much more efficient than STT. VIII. RELATED WORK Many research works suggest fast RFID identification schemes. In the ALOHA [31] scheme, tags reply with their IDs in random slots and only singleton slots can be identified. Their performances maximize when the issued frame size is optimized. Tree-based schemes [3], [13], [14], [16], [22] aim to reduce the number of collisions through reader-tag interactions. Hybrid approaches [18] combine both for identification purpose. Different from these works, we utilize physical layer information from tag replies to improve the identification efficiency. Collision recovery methods aim to decode collisions of multiple tags. Buzz [25] assumes bit-level synchronization among responding tags and decodes the aggregated rateless codes of multiple tags. [2], [23] analyze the constellation map and separate 2 tags. BiGroup [15] decodes up to 5-tag reply with reasonable accuracy by extracting temporal-spatial features from the received signal. Laissez-Faire [11] detects signal edges and separates signal edges of multiple tags and thereby decodes tag collisions. Laissez-Faire, however, requires the tags to transmit with assigned initial offsets and bit durations. Unlike those works, we focus on improving the overall query process to identify all tags with minimum number of queries. Existing works [9], [29] have explored how to utilize physical layer information to support other RFID operations. PLACE [9] counts the number of clusters in contellation map to enhance RFID cardinality estimation. Tagoram [29] utilizes phase information from RFID tags and tracks the labelled items. A recent ALOHA-based RFID identification work PUTI [33] is similar to PHY-Tree, as it also analyzes collided response slots to orientate the identification process. Nevertheless, it suffers from inherent issues in ALOHA-based works, i.e., tag starvation and unstable performance. IX. CONCLUSION Traditional tree-based RFID identification methods are on-ly MAC-layer solutions and perform poorly when the tag ID distribution is random. We propose a novel tree-based RFID identification scheme that utilizes physical layer information to improve the identification performance. Given the current queried tree node, PHY-Tree can skip its children nodes in either left or right branch by opportunistically detecting resonant states; meanwhile, it also estimates how many tags collide together and skips its children nodes in the same layers. By accumulating both information along the whole identification process, PHY-Tree significantly outperforms previous MAC layer tree-based schemes.

13 HOU AND ZHENG: PHY-TREE: PHYSICAL LAYER TREE-BASED RFID IDENTIFICATION 13 REFERENCES [1] (Jul. 2014). EPCglobal C1G2 UHF RFID Protocol at 860 MHz-960 MHz. [Online]. Available: org/standards/uhfc1g2 [2] C. Angerer, R. Langwieser, and M. Rupp, RFID reader receivers for physical layer collision recovery, IEEE Trans. Commun., vol. 58, no. 12, pp , Dec [3] J. Capetanakis, Tree algorithms for packet broadcast channels, IEEE Trans. Inf. Theory, vol. IT-25, no. 5, pp , Sep [4] B. Chen, Z. Zhou, and H. Yu, Understanding RFID counting protocols, IEEE/ACM Trans. Netw., vol. 24, no. 1, pp , Feb [5] N. Cho, S.-J. Song, S. Kim, S. Kim, and H.-J. Yoo, A 5.1-µw UHF RFID tag chip integrated with sensors for wireless environmental monitoring, in Proc. IEEE ESSCIRC, Sep. 2005, pp [6] J. Gan and Y. Tao, DBSCAN revisited: Mis-claim, un-fixability, and approximation, in Proc. ACM SIGMOD, 2015, pp [7] R. E. Greeff, F. W. Smith, and D. K. Ovard, RFID device time synchronization, U.S. Patent , Feb. 15, [8] H. Hassanieh, J. Wang, D. Katabi, and T. Kohno, Securing RFIDs by randomizing the modulation and channel, in Proc. USENIX NSDI, 2015, pp [9] Y. Hou, J. Ou, Y. Zheng, and M. Li, PLACE: Physical layer cardinality estimation for large-scale RFID systems, in Proc. IEEE INFOCOM, May 2015, pp [10] Y. Hou and Y. Zheng, PHY assisted tree-based RFID identification, in Proc. IEEE INFOCOM, May 2017, pp [11] P. Hu, P. Zhang, and D. Ganesan, Laissez-faire: Fully asymmetric backscatter communication, in Proc. ACM SIGCOMM, 2015, pp [12] R. S. Khasgiwale, R. U. Adyanthaya, and D. W. Engels, Extracting information from tag collisions, in Proc. IEEE RFID, Apr. 2009, pp [13] C. Law, K. Lee, and K.-Y. Siu, Efficient memoryless protocol for tag identification, in Proc. ACM DIALM Workshop, 2000, pp [14] V. Namboodiri and L. Gao, Energy-aware tag anti-collision protocols for RFID systems, in Proc. IEEE PerCom, Mar. 2007, pp [15] J. Ou, M. Li, and Y. Zheng, Come and be served: Parallel decoding for COTS RFID tags, in Proc. ACM MobiCom, 2015, pp [16] L. Pan and H. Wu, Smart trend-traversal: A low delay and energy tag arbitration protocol for large RFID systems, in Proc. IEEE INFOCOM, Apr. 2009, pp [17] J. Posamentier, RFID tag clock synchronization, U.S. Patent , Sep. 6, [18] C. Qian, Y. Liu, H. Ngan, and L. M. Ni, ASAP: Scalable identification and counting for contactless RFID systems, in Proc. IEEE ICDCS, Jun. 2010, pp [19] C. Qian, H. Ngan, and Y. Liu, Cardinality estimation for large-scale RFID systems, in Proc. IEEE PerCom, Mar. 2008, pp [20] J. Ryu, H. Lee, Y. Seok, T. Kwon, and Y. Choi, A hybrid query tree protocol for tag collision arbitration in RFID systems, in Proc. IEEE ICC, Jun. 2007, pp [21] M. Shahzad and A. X. Liu, Every bit counts: Fast and scalable RFID estimation, in Proc. ACM MobiCom, 2012, pp [22] M. Shahzad and A. Liu, Probabilistic optimal tree hopping for RFID identification, in Proc. ACM SIGMETRICS, 2013, pp [23] D. Shen, G. Woo, D. P. Reed, A. B. Lippman, and A. Lippman, Separation of multiple passive RFID signals using software defined radio, in Proc. IEEE RFID, Apr. 2009, pp [24] J. Wang, F. Adib, R. Knepper, D. Katabi, and D. Rus, RF-compass: Robot object manipulation using RFIDs, in Proc. ACM MobiCom, 2013, pp [25] J. Wang, H. Hassanieh, D. Katabi, and P. Indyk, Efficient and reliable low-power backscatter networks, in Proc. ACM SIGCOMM, 2012, pp [26] J. Wang and D. Katabi, Dude, where s my card?: RFID positioning that works with multipath and non-line of sight, in Proc. ACM SIGCOMM, 2013, pp [27] J. Wang, D. Vasisht, and D. Katabi, RF-IDraw: Virtual touch screen in the air using RF signals, in Proc. ACM SIGCOMM, 2014, pp [28] R. Want, An introduction to RFID technology, IEEE Pervas. Comput., vol. 5, no. 1, pp , Jan [29] L. Yang et al., Tagoram: Real-time tracking of mobile RFID tags to high precision using COTS devices, in Proc. ACM MobiCom, 2014, pp [30] R. Zhang, Y. Liu, Y. Zhang, and J. Sun, Fast identification of the missing tags in a large RFID system, in Proc. IEEE SECON, Jun. 2011, pp [31] B. Zhen, M. Kobayashi, and M. Shimizu, Framed ALOHA for multiple RFID objects identification, IEICE Trans. Commun., vol. 88, no. 3, pp , [32] Y. Zheng and M. Li, P-MTI: Physical-layer missing tag identification via compressive sensing, in Proc. IEEE INFOCOM, Apr. 2013, pp [33] F. Zhu, B. Xiao, J. Liu, and L. Chen, Efficient physical-layer unknown tag identification in large-scale RFID systems, IEEE Trans. Commun., vol. 65, no. 1, pp , Jan [34] A. Molisch, Wireless Communications, vol. 34. Hoboken, NJ, USA: Wiley, Yuxiao Hou received the B.S. degree in the special class for gifted young from the University of Science and Technology of China, China, in 2012, and the Ph.D. degree from the School of Computer Science and Engineering, Nanyang Technological University, Singapore, in His research interests include RFID, backscatter networks, and Internet of Things. Yuanqing Zheng received the B.S. degree in electrical engineering and the M.E. degree in communication and information system from Beijing Normal University, China, in 2007 and 2010, and the Ph.D. degree from Nanyang Technological University in He is currently an Assistant Professor with the Department of Computing, Hong Kong Polytechnic University. His research interest includes mobile and wireless computing, and RFID.