Research On Human Detection And Tracking Based On Deformable Part-based Model

Human detection and tracking is one of the key issues and research focuses in the field of computer vision. It combines pattern recognition, machine learning, image processing, artificial intelligence, and many other advanced technologies of several fields. Its main task is to detect and track humans in a continuous sequence of images, then get humans' trajectory. Human detection and tracking have broad application prospects and practical values in the fields of intelligent monitoring, automatic drive, intelligent robots, man-machine interaction. This paper makes several studies on human detection and tracking as follows:1. According to the relevant knowledges about the human's perception of color, we proposed the human detection algorithm of deformable part model based on opponent color space(OPP). PASCAL VOC 2007 dataset is utilized here to test the algorithm. Firstly, we transform all the images of PASCAL VOC 2007 dataset to OPP color space from RGB color space. Secondly, we train a classifier of deformable part model using the training dataset. Finally, testing dataset is tested by the trained classifier. Experimental results show that, compared to the detection of RGB color space, the proposed method which based on OPP color space not only enhances the detection precision, but also reduces the false alarm rate.2. For the disadvantages, the slow detection speed and high false alarm rate, of the deformable part-based model algorithm, we proposed a new human detection algorithm based on selected candidate detection positions, which combines deformable part-based model with the object proposals methods. The proposed method improves the original deformable part-based model algorithm by utilizing the merits(fast speed and pretty precise results) of object proposal methods. Firstly, we utilize the object proposal method to generate several positions that is considered probably include objects. Then, we detect objects only in these candidate positions, which avoid the exhaustive sliding window method. Experiments show that the proposed algorithm accelerates the detection speed, and reduces the false alarm rate.3. We combine the two aspects of object detection and tracking, proposed a tracking method which combines the Deformable Part-based Model and Meanshift algorithm. In addition, the proposed method utilizes Kalman Filter to improve the robustness of the tracking. Firstly, the proposed method takes advantage of deformable part-based model to detect the human. Then, we estimate the position of the human in next frame according to this frame. After that, Meanshift algorithm is utilized to search the best match. Experiments show that the proposed method improves the stability and robustness of the tracking.