Example-Based Color Vehicle Retrieval for Surveillance

In this paper, we evaluate several low dimensional color features for object retrieval in surveillance video. Previous work in object retrieval in surveillance has been hampered by issues in low resolution, poor segmentation, pose and lighting variations and the cost of retrieval. To overcome these difficulties, we restrict our analysis to alarm-based vehicle detection and as a consequence, we restrict both pose and lighting variations. In addition, we study the utility of example-based retrieval to avoid the limitations of strict color classification. Finally, since we perform our evaluation at run-time for alarm-based detection, we do not need to index into a large database. We evaluate the efficiency and effectiveness of several color features including standard color histograms, weighted color histograms, variable bin size color histograms and color correlograms. Results show color correlogram to have the best performance for our datasets.