Dynamic object construction using belief function theory

Several video surveillance applications aim at detecting the object(s) appeared in the scene of interest. Assuming these objects may be of any kind, change detection algorithms are often used to detect objects or object subparts, that have then to be gathered to form the objects. We call ‘object construction’ this step of object fragments (elementary change detections) collecting, associating to the right object and fusing with other fragments. In this work, we model the uncertainty and the imprecision of the location of the detected fragments using Belief Function Theory (BFT). During object construction, two mechanisms compete, namely the data accumulation and their temporal removal or weighting. We show that BFT framework is suitable for implementing these mechanisms as well as the data association between the new detections that are unlabeled and the objects under construction. Tests on actual data were performed. They allow for the quantitative evaluation of the proposed method in term of robustness versus the object partial occultation and crossing. Proposed approach is also compared with several alternative approaches.