Cooperative processing based on posture change detection and trajectory estimation for unknown multi-object tracking

Tracking of moving objects is a very important step in building an intelligent video surveillance system. The movement of non-rigid objects, appearance variations and luminosity changes make tracking even more difficult. This paper proposes a new automatic multi-target tracking system that can deal with the most confronted problems without any prior knowledge of the characteristics of objects. The system is a combination between classification, learning and tracking in a parallel architecture that allows the three tasks to be performed separately and efficiently to make the most of this combination. The permanent learning of the classifier guarantees the efficiency of the latter compared to the posture changes of moving objects. The classifier sends the new posture changes with a high degree of confidence as a new learning data. This cyclic aspect forces the system to adapt to all possible posture changes. In the case of occlusion, the system uses the estimated information of the trajectories to correct or cancel the learning process. The filtering process prevents the classifier from falling into a false classification, which significantly increases the system adaptability to the environment. Tests carried out on the CAVIAR and MOT16 datasets showed the efficiency and effectiveness of the proposed system.