Condensation-based multi-person detection and tracking with HOG and LBP

Multi-person tracking and detection is widely used in human robot interaction, which has been a hot topic in computer vision. In this paper, we utilize a tracking-by-detection framework to track many persons at the same time. We use HOG and LBP features to describe person´s characteristics in a scene and train a strong classifier using Adaboost algorithm. In the tracking part, we use a particle filter to estimate the targets´ position. Besides, we train an on-line SVM classifier to improve the accuracy of the tracking results by learning and updating the detector´s results. The particles´ velocity is also utilized to improve the accuracy of the data association, which relates the detector´s output to the tracker´s results. Our method is validated feasible on UBC-Hockey benchmark datasets.