Adaptive Credal Multi-Target Assignment for Conflict Resolving

This paper tackles the object assignment problem in a multiple target tracking context. Multi-target association consists in defining at each time step the relations between two sets. One set is related to the newly detected objects and the other to the already known ones. In this paper, the Dempster-Shafer theory, also known as belief functions is used. The contribution lies in the development of a methodology for cautious object association. Contradictions and ambiguities are managed using a single information matrix containing the assignment data. An algorithm allowing non-association in case of conflicting situations is evaluated and compared to state-of-the-art credal solutions. It particularly shows that in cluttered environment generating association conflict, the algorithm is able to postpone the pairing decision instead of propagating erroneous assignments. The robustness of the method is tested in different situations, including a traffic sign detection system. The optimization of the assignment method with respect to the desired risk level is shown.