Condensation-based Multi-person Tracking using an Online SVM Approach

We propose a multi-person tracking framework using only one single camera in this paper. We utilize particle filter as the tracking framework and train a SVM classifier by reliable examples extracted from associated detections without occlusion. Based on the results of data association, we integrate the target's velocity into weights calculation to handle object occlusion assuming that fast-moving target is not likely to change directions abruptly because of inertia. In addition, we design a new data association method whose affinity measure is computed by the classifier score judged on candidate image patch, the distance and size similarity of two rectangles. The experiments reveal that our method obtains a better performance compared with other state-of-the-art algorithms for PETS'09 videos S2 L1.