John Heidemann

Evaluating Signature Matching in a Multi-Sensor Vehicle Classification System (extended)

TitleEvaluating Signature Matching in a Multi-Sensor Vehicle Classification System (extended)
Publication TypeTechnical Report
Year of Publication2011
AuthorsC. Zhang, and J. Heidemann
Date Publishednov

Many academic sensornet systems consider the problem of tracking targets as they move through a fixed field of sensors. Such applications must relate sensor \emphdetections to specific \emphtargets, yet prior work has often ignored this problem, assuming either a single target, sufficient spatial separation that target-to-sensor mapping is clear, or some out-of-band detection-to-target mapping. There has been little study of algorithms to associate detections with targets, and effects of their accuracy on detection results, particularly in environments with dense targets. We explore the question of \emphdetection-to-target mapping in the context of a \emphvehicle classification system for urban roadways, where vehicles pass fixed sensors at varying but frequent rates. We develop several \emphsignature matching algorithms that relate detections at different sensors to same or different vehicles. We evaluate these algorithms with data taken in a field test of live traffic compared against ground truth obtained through manual analysis of video and the resulting matching recall is over 78%. We investigate the effects of mapping accuracy on length-based vehicle classification. We show that accurate signature matching is critical to multi-sensor algorithms. We compare our matching algorithms against an oracle (perfect information), and find that all matching reduces end-to-end accuracy somewhat, but a poor matching algorithm reduce accuracy by 21%, while our best algorithm reduces it by only 10% in our case study. Finally, we quantify the degree of correlation between matching correctness and classification accuracy.