A Comparative Evaluation of Well-known Feature Extractors for Multi-view Vehicle Tracking in VSNs

The choice of local features to use for feature-based vehicle tracking application is of great importance since the tracking accuracy depends directly on the robustness of the extracted features. Aside of the efficiency of features in terms of analysis accuracy, considering the computation speed and the required transmission bitrate in the choice of the feature extraction algorithm is crucial in order to design a system suitable for the constrained resources of visual sensor networks. In this paper, we evaluate the performance of the four most well-know algorithms for feature detection and extraction, SIFT, SURF, ORB, and BRISK, according to their contribution to the multi-view vehicle matching accuracy, to their computational speed, and to their demands in transmission bitrate. From the obtained results, we deduce that SIFT is the most accurate feature extractor for matching vehicle from two different views, but at the same time, is the most expensive algorithm. ORB is the fastest algorithm with lower matching accuracy compared to SIFT. BRISK represents a good compromise between the computation cost and the matching accuracy.