Multi-targets Tracking of Multiple Instance Boosting Combining with Particle Filtering

In order to surmount the major difficulties in multi-target tracking, one was that the observation model and target distribution was highly non-linear and non-Gaussian, the other was varying number of targets bring about overlapping complex interactions and ambiguities. We proposed a kind of system that is able of learning, detecting and tracking the multi-targets. In the method we combine the advantages of two algorithms: mixture particle filters and Multiple Instance Boosting. The key design issues in particle filtering are the selecting of the proposal distribution and the handling the problem of objects leaving and entering the scene. We construct the proposal distribution using a compound model that incorporates information from the dynamic models of each object and the detection hypotheses generated by Multiple Instance Boosting. The learned Multiple Instance Boosting proposal distribution makes us to detect quickly object which is entering the scene, while the filtering process allows us to keep the tracking of the simple object. An automatic multiple targets tracking system is constructed, and it can learn and detect and track the interest object. Finally, the algorithm is tested on multiple pedestrian objects in video sequences. The experiment results show that the algorithm can effectively track the targets the number is changed.