A novel multi-object tracking algorithm under occlusions

Multi-object tracking is one of challenging topics for Computer Vision. We describe a novel multi-object tracking algorithm based on cascade SVM for anti-occlusion. The classifer is divided into two levels. The first level(crude) classifer is to choose the most senitive blocks to reduce the number of negative samples for second level classifer. The second-level classifer fucus on these negative samples, increasing the correct classification rate of detection. For occluded objects, the new solution is to measure the similarity between objects. We esitblish the three lists to record the tracking information, including size, position, apparance and orientation of velocity.The low-level method is identified objects by these parameters. The high level method plays excellently on complex situation like one tracklet is occluded by others, which apply the estimation position for missing objects to caluate the similarity between them. The experiments demonstrate the accuracy rate of the algorithm.