Cleansing indoor RFID tracking data

The Radio-Frequency Identification (RFID) technology has been increasingly deployed in indoor environments for object tracking and monitoring. However, the uncertain characteristics of RFID data, including noise and incompleteness hinder RFID data querying and analysis at higher levels. Hence, it is of paramount importance to cleanse the RFID data for such applications. This paper introduces our comprehensive research on cleansing RFID data in indoor settings. We focus on two inherent errors in such RFID data: false positives (unexpected cross readings) and false negatives (missing readings). In our proposed graph model based approach, we design a probabilistic distance-aware graph to represent the indoor topology, the deployment of RFID readers and their sensing parameters. We also augment the graph with transition probabilities that capture how likely objects move from one RFID reader to another. Based on the proposed graph, we design cleansing algorithms to reduce false positives and recover false negatives. In the learning-based approach, we propose an Indoor RFID Multi-variate Hidden Markov Model (IR-MHMM) to capture the uncertainties of indoor RFID data as well as the correlation of moving object locations and object's RFID readings. We solely use raw RFID data for the learning of the IR-MHMM parameters. Using the resulting IR-MHMM, the learning-based approach is able to deliver cleansing performance comparable to and even better than that of the graph model based approach, although the former requires much less prior knowledge than the latter.