Research On Pedestrian Single Target Tracking Algorithm Based On Discriminative Model

Pedestrian target tracking technology is an important research area in computer vision,which has important applications in unmanned driving,intelligent surveillance,intelligent transportation,etc.,and has received wide attention in recent years.Although many tracking methods have been proposed at home and abroad,due to the complexity of tracking scenes with occlusions,background similarities interference,and targets out of the field of view,the existing algorithms are difficult to achieve effective tracking for pedestrian targets in some complex scenes.In this paper,based on the deep learning discriminative online update tracking algorithm model,we conduct an in-depth analysis of experiments,actively explore and propose an improvement scheme for the drift,loss and tracking inaccuracy caused by background interference,occlusion between pedestrians and out-of-field conditions encountered by the tracker in the pedestrian tracking process,so as to improve the accuracy and robustness of the tracking model.The main research work of this paper is as follows.(1)A deep discriminative pedestrian single-target tracking algorithm incorporating pedestrian characteristics is proposed based on a discriminative online updated tracking model to address the problems of drift and tracking inaccuracy brought by the tracking model due to the problems of occlusion and interference of similar objects in the background.In the classification task,excitation and suppression losses are proposed to improve the discriminative ability of the classifier for target and background similarities;different tracking states are defined in different scenes according to the complexity of the current scene,and different online update strategies are adopted for the classification filters to increase the discriminative ability of the classifier;the coarse position of the target is corrected by incorporating the inherent characteristics of the pedestrian movement process.The error accumulation is reduced by adding candidate bounding boxes in the regression task according to the pedestrian characteristics.The above improvements enable the tracking model to effectively discriminate between the target and the background during the tracking process,and to maintain robustness and efficiency in the tracking process.(2)A cross-view scene matching pedestrian single target tracking algorithm is proposed for the target loss problem caused by pedestrians blocking each other and out of view in complex scenes.The cross-view spatial matching is performed by introducing the information of the top view in the same scene.The spatial positions of the same pedestrian target in horizontal view and top view are matched,and then combined with the tracking information of the same target in the video sequence of top view,effective predictive tracking of the target pedestrian in horizontal view is performed under the guidance of the spatial information of the pedestrian target in top view.The above proposed tracking method can effectively solve the problem of pedestrian occlusion and out of view in part of the horizontal view.