Research On Vehicle Tracking Algorithm Based On Adaptive Visual Features

Vehicle tracking algorithm is a branch of object tracking field,which limits tracking object to vehicle.It means to determine the positions of the same vehicle in different frame images and connect them into a trajectory based on computer vision technology.The applications based on this technology include obtaining vehicle speed,detecting the occurrence of traffic accidents,tracking suspected vehicles,etc.Therefore,vehicle tracking technology is widely used in intelligent transportation,smart city,human-computer interaction,and other fields.The research in the field of object tracking has made some achievements.However,the existence of occlusion,illumination change,angle change,distant object and other factors still affect the effectiveness of the tracking algorithm.Aiming at the problems of dense vehicle flow,similar vehicle appearance and changeable scenes,a robust detection-based vehicle tracking algorithm is proposed in this paper.The research contents and innovation achievements can be summarized as follows:(1)For the problem of object occlusion caused by dense traffic,Mask-RCNN trained on the MS COCO dataset is used as the detector to extract the features of the mask.And the vehicle type is used as one of the features,which is modified by IOU to optimize the output of the detector.(2)For the problem of unstable algorithm effect caused by the changeable application scenes,decision tree algorithm is used to fuse multiple visual features adaptively.The experiments prove that the above method can effectively enhance the robustness of the algorithm after comparing three feature fusion methods:fixed weight,standard deviation weight,and decision tree weight.(3)For the problem of ID switch during vehicles with similar appearances approaching,5 kinds of tracklets matching features are proposed in this paper,which enriches the basis of object matching.And the tracklets clustering operation is performed based on all features.The proposed method is proved to be able to improve the tracking accuracy through comparative experiments.Experiments on two public data sets show that the proposed algorithm performs well in overall accuracy and is stable in different scenes.