Online Multiple Pedestrians Tracking Based On Detection And Feature Extraction Network

Online multi-target tracking has important applications in many video analysis scenar-ios,such as robot navigation and automatic drive.One of the most important issues in online multi-target tracking is how to improve the speed of the algorithm to meet the needs of many real-time scenes.Current tracking methods usually do target detection in each frame first and then follow the result of detection.The corresponding image features are often extracted for each target detected and the final trajectories of each target are obtained according to some matching algorithms or graph algorithms.But these methods ignore the fact that in the procedure of object detection,image feature of targets already exists in the deep neural network,but the existing methods are only use the bounding box,feature is extracted by an-other neural network or hand-craft methods.Secondly,the existing methods in the process of target detection is detection algorithm directly uses multi classification,but the detection of pedestrian is different from the ordinary multi-class detection algorithm,such as the mutual occlusion of two targets of the same class.in this case the target detection algorithm will be unable to distinguish two object's accurate boundary.Finally;in the process of matching targets,most of them will take the motion information of the target into account,and use.the distance between objects to judge the similarity between them.A common method of motion estimation is the Kalman filter.Kalman filter can be used to obtain smoother target trajectory,but the prediction of the target position by the Kalman filter will become inac-curate when the target is occluded for a long time.In this paper,a method based on block matching is proposed to reduce the region that the target may appear.In general,the main contributions of this article are as follows:1.A network that simultaneously obtains the target detection results and the image fea-tures corresponding to each target is designed.The image feature of the target becomes a "by-product" of the target detection process,and the subsequent target tracking pro-cess is obtained.A large degree of simplification requires only the similarity between the features of the image as a measure of the use of a matching algorithm the trajectory of each target.2.To locate target after long time occlusion better,a block matching method is proposed to further reduce the location of the target and reduce the target matching error after long time occlusion.