A probabilistic approach to address uncertainty of RFID

The popularity of Radio Frequency Identification (RFID) has increased in recent years. Although RFID has many benefits over other identification technologies, there are certain situations that can lead to uncertainty about the true location of the tagged object. In this paper we incorporate the knowledge about the operation of an RFID system and its various failure modes, and evidence from the RFID reader such as low-level diagnostic information in a single probabilistic model. The probabilistic model allows us to quantify and reduce the uncertainty associated with RFID operation, since we can provide the application with a single probability value that expresses the belief that a certain tagged object is present. The probabilistic framework is demonstrated with the help of a smart medicine cabinet. Radio Frequency Identification (RFID) systems have become common in places where access control and tracking of physical objects is required. Examples include cattle herding, car immobilizers, and transport ticketing [3]. More recently, RFID systems have begun to find greater use in the consumer object identification market, in industrial automation, and in supply chain management. The use of RFID systems in these application domains has been promoted by efforts to develop low cost RFID tags as an economical replacement of barcodes [5]. In the ubiquitous computing research community the potential of RFID tags has also been demonstrated by many researchers over the years. Examples include prototypes such as the Magic Medicine Cabinet by Wan [6], the augmentation of desktop items by Want et al. [7] and smart shelves by Decker et al. [2]. RFID systems consist of radio frequency (RF) tags, tag readers, and some software to process the tag reads. The tags typically respond to an RF broadcast by the tag reader by sending their serial number or other data stored in their memory to the reader. Compared to optical bar code systems, RFID tags have the advantage that they can be read without line of sight through non-conducting materials and that multiple tags can be detected nearly simultaneously. As such, RFID systems are a useful tool in tracking the location of physical objects. However, due to the low-cost and power constraints of RFID tags, reliability concerns arise under certain circumstances. In particular, two types of undesirable effects have been noticed [1]: – False negative reads, where RFID tags might not be read at all, leading to the mistaken belief that the object is not present, and, – False positive reads, where RFID tags might be read when they are outside the region normally associated with the location of the RFID reader, leading to a mistaken belief that the object is present. While some errors are of continuous nature, e.g. tags which are attached to metal surface might never be detected, most false negative and positive reads only occur temporarily, e.g. collisions on the air interface or interference from other radio sources. For a location application that uses RFID to judge whether a number of tagged objects are present, e.g. a smart shelf, the temporary false negative and positive reads result in unreliable operation. A key to reduce the uncertainty associated with RFID is fusing the low-level diagnostics information available from an RFID reader, with knowledge about how RFID works and contextual information on the objects equipped with RFID tags. In this paper we present a probabilistic framework that accounts for the probability distribution of various error sources, diagnostics information and contextual information. Given evidence from the reader and the probabilistic model of RFID operation it allows us to infer more accurately whether a tag is present at a certain location. Although this approach has applications in many domains where RFID is used, the motivating scenarios include those where multiple tags are present simultaneously and interfere with each other and where the application needs to provide instant feedback to the user on the objects which are present. Our approach is built on earlier work on Bayesian networks, which are known to perform well in domains containing uncertainty [4]. In the RFID domain, the uncertainty is due to incomplete knowledge of the state of the domain at the time, when the identification is performed, and the randomness in the mechanisms governing the behavior of the domain. The modular structure of the Bayesian network also lends itself to the inclusion of contextual information on the objects being tagged. This paper also shows how a dedicated tag that is attached to the reader antenna can be integrated into our probabilistic framework and how the reliability of operation is significantly increased.