Estimating Data Stream Quality for Object-Detection Applications

Abstract

Object-detection applications rely on streams of data gathered from sensors, RFID readers, and image recognition systems, among others. These raw data streams tend to be noisy, including both false positives (erroneous readings) and false negatives (missed readings). Techniques exist for general-purpose {\em cleaning} of these types of data streams, based on temporal and/or spatial correlations, as well as properties of the physical world. Cleaning is effective at improving the quality of the data, however no cleaning procedures can eliminate all errors. In this paper we identify and address the problem of {\em quality estimation} as object-detection data streams are cleaned. We provide techniques for estimating both {\em confidence} and {\em coverage} as streams are processed by cleaning modules. Detailed experimental results based on an RFID application demonstrate the accuracy and effectiveness of our approach.