Research On Road Vehicle Tracking Algorithm Based On Depth Learning

Due to many traffic problems caused by the sharp increase in the number of vehicles in recent years,relevant personnel proposed the concept of "intelligent transportation system".As the core link,vehicle tracking can automatically obtain the type,position and trajectory of vehicles in the road environment,which is very important for the decision-making and planning task of "intelligent transportation system",so the vehicle tracking algorithm has important research significance.This paper proposes a road vehicle tracking algorithm based on deep learning.The specific research contents are as follows:For the vehicle detection link in multi vehicle tracking,this paper proposes an improved vehicle detection network based on resnet-34 network.Specifically,the deep layer aggregation(DLA)is combined with resnet-34 network to improve the recognition ability of multi-scale vehicle targets;Combined with deformable convolution(DCN),it can reduce the influence of deformation on target feature extraction and improve the recognition ability of the network for different types of vehicles;Then local importance based pooling(LIP)is applied to replace the traditional down sampling pooling method,so that the network can more accurately distinguish the foreground target features and background features in the image,so as to reduce the impact of background occlusion on target detection.In the experimental stage,the vehicle detection network proposed in this paper is compared longitudinally.The experimental results show that compared with the improved resnet-34 network,the improved detection network has higher detection accuracy.For the multi vehicle tracking problem,this paper first combines the vehicle identity feature extraction branch with the improved vehicle detection network according to the idea of joint detection and tracking algorithm.Secondly,for the data association stage of multi vehicle tracking,this paper proposes a new data association algorithm based on Kalman filter and Hungarian matching algorithm,which dynamically determines the initial threshold of the new trajectory by dividing the detection results into two different sets,so that the algorithm can adapt to the changes of the detection environment.After that,by analyzing the characteristics of multi vehicle tracking scene,a new matching strategy is designed in the secondary data association to make up for the errors in the primary matching,so that the data association algorithm can make full use of the low score detection results to reduce the number of target ID switching in the process of data Association.Finally,the network is trained with the detrac dataset and the visualization results are displayed.In addition,the Kitti dataset is used to conduct horizontal comparison experiments between multiple algorithms.The results show that the multi vehicle tracking algorithm proposed in this paper effectively improves the tracking effect.