Citation Context

...indow size for each tag i, based on guidance from its binomial-sampling model as discussed above.4 Algorithm 1 SMURF Adaptive Per-Tag Cleaning Require: T = set of all observed tag IDs δ = required completeness confidence ∀i ∈ T, wi ← 1 while (getNextEpoch()) do for (i in T ) do processWindow(Wi) w∗i ← completeSize(p avg i , δ) // Lemma 4.1 if (w∗i > wi) then wi ←max{min{wi + 2, w∗i }, 1} else if (detectTransition(|Si|, wi, pavgi )) then wi ←max{min{wi/2, w∗i }, 1} end if end for end while 3More conservative, non-CLT-based probabilistic criteria, e.g., based on the Chebyshev or Chernoff bounds [25] can also be used here. 4Note that our algorithm uses only simple mathematical operations and, thus, the overhead beyond traditional smoothing is minimal. SMURF runs a sliding-window aggregate for each observed tag i. The window size is initially set to one epoch for each tag, and then adjusted dynamically based on observed readings. (If at any point during processing SMURF sees an empty window for a tag, it resets its window size to one epoch.) During each new epoch, and for each tag i, SMURF starts by processing the readings of tag i inside the window Wi (processWindow(Wi)). This processing ...

Citation Context

...ID unreliability by modeling observed RFID readings as an unequal-probability random sample of tags in the physical world. This approach allows SMURF to balance the tension between reader unreliability and tag dynamics in a principled, statistical manner by continuously adapting the smoothing strategy to provide accurate, unbiased data to applications. (Section 3) • An Adaptive Smoothing Filter for RFID Data. Building on SMURF’s sampling-based foundation, we propose two novel, adaptive smoothing mechanisms for (a) cleaning the readings of single tag using techniques based on binomial sampling [12] (per-tag cleaning), and (b) cleaning an aggregate signal (e.g., count) over a tag population based on π- (or Horvitz-Thompson) estimators [29] (multi-tag cleaning). (Section 4) • An Experimental Study Validating the Effectiveness of SMURF’s Cleaning Algorithms. We present a detailed experimental study using various schemes to clean both synthetic and real RFID data streams. First, these tests show that there is no single static window size that works well in all environments (reader and tag behavior), motivating the need for an adaptive approach. Second, we demonstrate SMURF’s ability to adap...

Citation Context

...tag cleaning). Additionally, SMURF incorporates two modules that apply to both data-cleaning techniques: a sliding-window processor for fine-grained RFID data smoothing, and an optimization mechanism for improving cleaning effectiveness by detecting mobile tags. We first briefly discuss how SMURF processes readings within its adaptive window, then detail the two key cleaning mechanisms used by SMURF, and finally present SMURF’s mobile tag detection enhancement. 4.1 SMURF Sliding Window Processing Window-based smoothing in SMURF closely resembles traditional sliding-window aggregate processing [2, 7, 10] as expressed, for example, in Query 1 (of course, with the fixed-size Range clause removed). Similar to other RFID smoothing filters, SMURF produces a tag reading for a window if there exists at least one reading for the tag within that window [20, 24]. To enable our more sophisticated data-cleaning schemes, SMURF’s sliding-window processor also implements two basic modifications to conventional RFID filters: (1) partitioned RFID smoothing, and (2) epoch-based mid-window slide. As subsets of tagged objects may behave very differently (e.g., in a warehouse environment, some tagged items may be...

Citation Context

... in the current window, the value of |Si |is within±2 p Var[|Si|] of its expectation with probability close to 0.98. Based on this observation, SMURF flags a transition (i.e., exit) for tag i in the current window if the number of observed readings is less than the expected number of readings and the following condition holds:3 ||Si |− wipavgi |> 2 · q wip avg i (1− p avg i ) (1) SMURF Per-Tag Cleaning Algorithm. A pseudo-code description of SMURF’s adaptive per-tag cleaning algorithm is depicted in Algorithm 1. SMURF employs the common AdditiveIncrease/Multiplicative-Decrease (AIMD) paradigm [11] to adjust its window size for each tag i, based on guidance from its binomial-sampling model as discussed above.4 Algorithm 1 SMURF Adaptive Per-Tag Cleaning Require: T = set of all observed tag IDs δ = required completeness confidence ∀i ∈ T, wi ← 1 while (getNextEpoch()) do for (i in T ) do processWindow(Wi) w∗i ← completeSize(p avg i , δ) // Lemma 4.1 if (w∗i > wi) then wi ←max{min{wi + 2, w∗i }, 1} else if (detectTransition(|Si|, wi, pavgi )) then wi ←max{min{wi/2, w∗i }, 1} end if end for end while 3More conservative, non-CLT-based probabilistic criteria, e.g., based on the Chebyshev or ...

Citation Context

... SMURF to balance the tension between reader unreliability and tag dynamics in a principled, statistical manner by continuously adapting the smoothing strategy to provide accurate, unbiased data to applications. (Section 3) • An Adaptive Smoothing Filter for RFID Data. Building on SMURF’s sampling-based foundation, we propose two novel, adaptive smoothing mechanisms for (a) cleaning the readings of single tag using techniques based on binomial sampling [12] (per-tag cleaning), and (b) cleaning an aggregate signal (e.g., count) over a tag population based on π- (or Horvitz-Thompson) estimators [29] (multi-tag cleaning). (Section 4) • An Experimental Study Validating the Effectiveness of SMURF’s Cleaning Algorithms. We present a detailed experimental study using various schemes to clean both synthetic and real RFID data streams. First, these tests show that there is no single static window size that works well in all environments (reader and tag behavior), motivating the need for an adaptive approach. Second, we demonstrate SMURF’s ability to adapt its data-cleaning strategy to a wide range of reader characteristics and tag behaviors; in an environment with changing conditions, SMURF red...

Citation Context

...tag cleaning). Additionally, SMURF incorporates two modules that apply to both data-cleaning techniques: a sliding-window processor for fine-grained RFID data smoothing, and an optimization mechanism for improving cleaning effectiveness by detecting mobile tags. We first briefly discuss how SMURF processes readings within its adaptive window, then detail the two key cleaning mechanisms used by SMURF, and finally present SMURF’s mobile tag detection enhancement. 4.1 SMURF Sliding Window Processing Window-based smoothing in SMURF closely resembles traditional sliding-window aggregate processing [2, 7, 10] as expressed, for example, in Query 1 (of course, with the fixed-size Range clause removed). Similar to other RFID smoothing filters, SMURF produces a tag reading for a window if there exists at least one reading for the tag within that window [20, 24]. To enable our more sophisticated data-cleaning schemes, SMURF’s sliding-window processor also implements two basic modifications to conventional RFID filters: (1) partitioned RFID smoothing, and (2) epoch-based mid-window slide. As subsets of tagged objects may behave very differently (e.g., in a warehouse environment, some tagged items may be...

Citation Context

...a. Several projects have explored simple techniques to clean RFID data, typically based on fixed-window smoothing. In one paper, the authors identify the trade-off between smoothing the data and capturing the temporal variation but provide no real solutions [18]. In previous work, we recognize the need to clean RFID data and use an approach based on declarative continuous queries [19, 21, 22]. We show smoothed RFID data using different sized windows, but do not address how to choose the best size. The idea of using probabilistic models for sensor measurements has been explored in earlier work [15]; still, our work is the first to apply statistical techniques for adaptive RFID data cleaning. Furthermore, this scheme relies on learning and maintaining many fairly heavyweight multi-dimensional Gaussian models; our techniques rely on simple, non-parametric sampling estimators. Exploring interactions between the two approaches is an interesting area for future work. Adaptive filtering has been studied in digital signal processing in wide-ranging contexts such as image analysis and speech processing [27]. Especially applicable are nonlinear digital filters, which are designed to capture tran...

Citation Context

... of using probabilistic models for sensor measurements has been explored in earlier work [15]; still, our work is the first to apply statistical techniques for adaptive RFID data cleaning. Furthermore, this scheme relies on learning and maintaining many fairly heavyweight multi-dimensional Gaussian models; our techniques rely on simple, non-parametric sampling estimators. Exploring interactions between the two approaches is an interesting area for future work. Adaptive filtering has been studied in digital signal processing in wide-ranging contexts such as image analysis and speech processing [27]. Especially applicable are nonlinear digital filters, which are designed to capture transitions in the signal. For instance, AWED [27] adapts the size of a smoothing window for cleaning noisy images using a multi-phase approach involving smoothing and edge detection that inspired the basic SMURF design. 7. CONCLUSIONS AND FUTURE WORK While RFID technology holds much promise, the unreliability of the data produced by RFID readers is a major factor hindering large-scale deployment. Specifically, RFID readers suffer from low read rates, frequently failing to read tags that are present. Current s...

Citation Context

...e incorporated into an RFID middleware system to provide self-tuning smoothing without requiring the application to set this parameter. As a result, these systems become simpler to deploy and produce more reliable data. Several projects have explored simple techniques to clean RFID data, typically based on fixed-window smoothing. In one paper, the authors identify the trade-off between smoothing the data and capturing the temporal variation but provide no real solutions [18]. In previous work, we recognize the need to clean RFID data and use an approach based on declarative continuous queries [19, 21, 22]. We show smoothed RFID data using different sized windows, but do not address how to choose the best size. The idea of using probabilistic models for sensor measurements has been explored in earlier work [15]; still, our work is the first to apply statistical techniques for adaptive RFID data cleaning. Furthermore, this scheme relies on learning and maintaining many fairly heavyweight multi-dimensional Gaussian models; our techniques rely on simple, non-parametric sampling estimators. Exploring interactions between the two approaches is an interesting area for future work. Adaptive filtering ...

Citation Context

...tag cleaning). Additionally, SMURF incorporates two modules that apply to both data-cleaning techniques: a sliding-window processor for fine-grained RFID data smoothing, and an optimization mechanism for improving cleaning effectiveness by detecting mobile tags. We first briefly discuss how SMURF processes readings within its adaptive window, then detail the two key cleaning mechanisms used by SMURF, and finally present SMURF’s mobile tag detection enhancement. 4.1 SMURF Sliding Window Processing Window-based smoothing in SMURF closely resembles traditional sliding-window aggregate processing [2, 7, 10] as expressed, for example, in Query 1 (of course, with the fixed-size Range clause removed). Similar to other RFID smoothing filters, SMURF produces a tag reading for a window if there exists at least one reading for the tag within that window [20, 24]. To enable our more sophisticated data-cleaning schemes, SMURF’s sliding-window processor also implements two basic modifications to conventional RFID filters: (1) partitioned RFID smoothing, and (2) epoch-based mid-window slide. As subsets of tagged objects may behave very differently (e.g., in a warehouse environment, some tagged items may be...

Citation Context

... multiple reader interrogation cycles are typically grouped into what we term epochs.1 An epoch may be specified as a number of interrogation cycles or as a unit of time. A typical epoch range is 0.2-0.25 seconds [1, 31]. For each epoch, the reader keeps track of all the tags it has identified, as well as additional information such as the number of interrogation responses for each tag and the time at which the tag was last read. Readers store this information internally in a tag list (Table 1) which is periodically transferred to readers’ clients. For more information on RFID technology, see [34]. Tag ID Responses Timestamp 8576 2387 2345 8678 9 11:07:05 8576 4577 3467 2357 1 11:07:05 8576 3246 3267 5685 7 11:07:06 Table 1: Example reader tag list. RFID Reader and Tag Performance. To better understand the unreliability of RFID readings, we profile two RFID readers with different tags in two environments. Our profiling methodology is as follows. We suspend a single tag at varying distances in the same plane as the antenna. For every 6-inch increment of distance from the reader, we measure the read rate (number of responses to number of interrogations) for 100 epochs. 1In ALE terms, an ...

Citation Context

...in a principled manner to provide accurate RFID data to applications. 1. INTRODUCTION RFID (Radio Frequency IDentification) technology promises revolutions in areas such as supply chain management and ubiquitous computing enabled by pervasive, low-cost sensing and identification [17]. One of the primary factors limiting the widespread adoption of RFID technology is the unreliability of the data streams produced by RFID readers [8, 23]. The observed read rate (i.e., percentage of tags in a reader’s vicinity that are actually reported) in real-world RFID deployments is often in the 60−70% range [21, 23]; in other words, over 30% of the tag readings are routinely dropped. Unfortunately, such error rates render raw RFID streams essentially useless for the purposes of higher-level applications (such as accurate inventory tracking). Instead, RFID middleware systems are typically deployed between the readers and the application(s) in order to correct for dropped readings and provide “clean” RFID readings to application logic. The standard data-cleaning mechanism in most such systems is a temporal “smoothing filter”: a sliding window over the reader’s data stream that interpolates for lost reading...

Citation Context

... in reaction to the tags’ movement while using π- estimators to avoid under-counting with such a small window. As can be seen, there is no single static window that the warehouse monitoring application case use to provide accurate counts in this scenario. Using SMURF, in contrast, the application can get accurate readings throughout the pallet’s lifetime without setting the smoothing window size. π-SMURF further refines its accuracy by providing an unbiased estimate. 6. RELATED WORK Many commercial RFID middleware solutions contain configurable filters to process data produced by RFID readers [9, 20, 24, 33]. Many of these platforms explicitly incorporate data smoothing as a solution to RFID unreliability. None of these systems, however, provide any guidance for setting the size of the smoothing window. SMURF is designed to be incorporated into an RFID middleware system to provide self-tuning smoothing without requiring the application to set this parameter. As a result, these systems become simpler to deploy and produce more reliable data. Several projects have explored simple techniques to clean RFID data, typically based on fixed-window smoothing. In one paper, the authors identify the trade-o...

Citation Context

...field). – Completeness: To ensure that all tags in the reader’s detection range are read, the smoothing window must be large enough to correct for reader unreliability. Small window sizes cause readings for some tags to be lost, leading to false negatives (i.e., tags mistakenly assumed to have exited the reader’s detection range) and, consequently, a large underestimation bias (e.g., always under-counting the tag population). Adjusting the window size for completeness depends on the reader’s read rate, which, in turn, depends on both the type of reader and tag as well as physical surroundings [13, 18]. – Tag Dynamics: Using a large smoothing window, on the other hand, risks not accurately detecting tag movements within the window, leading to false positives (i.e., tags mistakenly assumed to be present after they have exited the reader’s detection range). Adjusting the window size for tag dynamics depends on the movement characteristics of the tags, which, in turn, can vary significantly depending on the application; for instance, a tag sitting on a shelf exhibits a different movement pattern from a tag on a conveyor belt. 163 Any RFID deployment must seriously consider and study the factor...

Citation Context

...eadings are routinely dropped. Unfortunately, such error rates render raw RFID streams essentially useless for the purposes of higher-level applications (such as accurate inventory tracking). Instead, RFID middleware systems are typically deployed between the readers and the application(s) in order to correct for dropped readings and provide “clean” RFID readings to application logic. The standard data-cleaning mechanism in most such systems is a temporal “smoothing filter”: a sliding window over the reader’s data stream that interpolates for lost readings from each tag within the time window [20, 24]. The goal, of course, is to reduce or eliminate dropped readings by giving each tag more opportunities to be read within the smoothing window. While the APIs for RFID middleware systems vary, smoothing filter functionality can be expressed as a simplified continuous query ∗This work was funded in part by NSF under ITR grants IIS-0086057 and SI-0122599, and by research funds from Intel and the UC MICRO program. Permission to copy without fee all or part of this material is granted provided that the copies are not made or distributed for direct commercial advantage, the VLDB copyright notice an...

Citation Context

...cts the results from two different profiling experiments that are representative of the 8 different profiles we collected. (The plots show the read rate of the tag at distances ranging from 0 to 20 feet.) All of the profiles have similar properties despite being generated using different readers, tags, and environments. First, the overall detection range of all readers and tags profiled remains relatively constant at 15-20 feet. Second, within each reader’s detection range, there are two distinct regions: (1) The area directly in front of the reader, termed the reader’s major detection region [26], giving high detection probabilities (read rates at or above 95%); and, (2) the reader’s minor detection region, extending from the end of the major detection region to the edge of the reader’s full detection range, where the read rate drops off linearly (with some variation) to zero at the end of the detection range. The main difference between our observed profiles lies in the percentage of the reader’s detection range corresponding to its major detection region. For instance, the major detection region corresponds to roughly 75% of the full detection range for the profile in Figure 2(a), w...

Citation Context

...linearly (with some variation) to zero at the end of the detection range. The main difference between our observed profiles lies in the percentage of the reader’s detection range corresponding to its major detection region. For instance, the major detection region corresponds to roughly 75% of the full detection range for the profile in Figure 2(a), whereas it makes up only 25% of the range in the profile in Figure 2(b). Note that our profiles are consistent with the results of in-depth commercial studies of the performance of many different tags and readers under highly-controlled conditions [14]. We also profile the readers to determine how they respond to the presence of multiple tags in their detection ranges. For these tests (not shown here), we suspend 10 tags in the same plane as the reader and measure the average read rate for 100 epochs at varying distances from the reader. While the overall properties of the observed profile does not change (we still find a separation between a major and minor detection region), the read rate in the major detection region typically drops to around 80%. Additional tests show that the read rate in the major detection region stays somewhat const...

Citation Context

...ll environments (reader and tag behavior), motivating the need for an adaptive approach. Second, we demonstrate SMURF’s ability to adapt its data-cleaning strategy to a wide range of reader characteristics and tag behaviors; in an environment with changing conditions, SMURF reduces overall error by a factor of more than 3 compared to the best environment-specific static window. (Section 5) SMURF is designed to be a component in a pipeline of operators responsible for low-level RFID data processing tasks such as cleaning, filtering, and spatial processing (see proposals such as ALE [5] and ESP [21, 22]). SMURF would be responsible for smoothing RFID readings from each reader before the streams are sent to other modules for additional processing. In this work, we focus on cleaning readings from a single reader or collection of logically equivalent readers (i.e., a logical reader [5]). We consider cleaning using multiple readers in Section 7 as ongoing work. SMURF’s sampling-based foundation offers a powerful conceptual framework for effective RFID data-cleaning tools. The set of techniques proposed in this paper can be directly incorporated in RFID middleware platforms to yield systems that ...

Citation Context

...ampling theory to drive its cleaning processes. Through the use of tools such as binomial sampling and π-estimators, SMURF continuously adapts the smoothing window size in a principled manner to provide accurate RFID data to applications. 1. INTRODUCTION RFID (Radio Frequency IDentification) technology promises revolutions in areas such as supply chain management and ubiquitous computing enabled by pervasive, low-cost sensing and identification [17]. One of the primary factors limiting the widespread adoption of RFID technology is the unreliability of the data streams produced by RFID readers [8, 23]. The observed read rate (i.e., percentage of tags in a reader’s vicinity that are actually reported) in real-world RFID deployments is often in the 60−70% range [21, 23]; in other words, over 30% of the tag readings are routinely dropped. Unfortunately, such error rates render raw RFID streams essentially useless for the purposes of higher-level applications (such as accurate inventory tracking). Instead, RFID middleware systems are typically deployed between the readers and the application(s) in order to correct for dropped readings and provide “clean” RFID readings to application logic. The...

Citation Context

...ampling theory to drive its cleaning processes. Through the use of tools such as binomial sampling and π-estimators, SMURF continuously adapts the smoothing window size in a principled manner to provide accurate RFID data to applications. 1. INTRODUCTION RFID (Radio Frequency IDentification) technology promises revolutions in areas such as supply chain management and ubiquitous computing enabled by pervasive, low-cost sensing and identification [17]. One of the primary factors limiting the widespread adoption of RFID technology is the unreliability of the data streams produced by RFID readers [8, 23]. The observed read rate (i.e., percentage of tags in a reader’s vicinity that are actually reported) in real-world RFID deployments is often in the 60−70% range [21, 23]; in other words, over 30% of the tag readings are routinely dropped. Unfortunately, such error rates render raw RFID streams essentially useless for the purposes of higher-level applications (such as accurate inventory tracking). Instead, RFID middleware systems are typically deployed between the readers and the application(s) in order to correct for dropped readings and provide “clean” RFID readings to application logic. The...

Citation Context

...eadings produced by a tag passing through an RFID-enabled door can be generated by moving a tag from outside DetectionRange to directly in front of the reader, and then back to outside DetectionRange. Tags move between 0 and 20 feet following one of two behaviors representative of a range of RFID applications: 1. Pallet: All tags have the same velocity. This simulates grouped tags, such as tagged items on a pallet. 2. Fido: Each tag chooses a random initial velocity (uniform between 1 and 3 feet/epoch). Note that the average velocity, 2 feet/epoch, is roughly equivalent to conveyor-belt speed [28]. Every 100 epochs, on average, each tag switches from a moving state to a resting state (and vice versa). When a tag resumes movement, it chooses another random velocity between 1 and 3 feet/epoch. This behavior simulates tracking environments such as a digital home, where each tag displays independent random behavior. Data Generation. We run the generator for NumEpochs epochs.7 At each epoch, the generator determines which tags are detected based on the read rate at each tag’s location relative to the reader. It then produces a set of readings containing a tag ID, epoch number, and the tag’s...

Citation Context

...field). – Completeness: To ensure that all tags in the reader’s detection range are read, the smoothing window must be large enough to correct for reader unreliability. Small window sizes cause readings for some tags to be lost, leading to false negatives (i.e., tags mistakenly assumed to have exited the reader’s detection range) and, consequently, a large underestimation bias (e.g., always under-counting the tag population). Adjusting the window size for completeness depends on the reader’s read rate, which, in turn, depends on both the type of reader and tag as well as physical surroundings [13, 18]. – Tag Dynamics: Using a large smoothing window, on the other hand, risks not accurately detecting tag movements within the window, leading to false positives (i.e., tags mistakenly assumed to be present after they have exited the reader’s detection range). Adjusting the window size for tag dynamics depends on the movement characteristics of the tags, which, in turn, can vary significantly depending on the application; for instance, a tag sitting on a shelf exhibits a different movement pattern from a tag on a conveyor belt. 163 Any RFID deployment must seriously consider and study the factor...