Research On Multi-scale Vehicle Target Detection And Tracking In Freeway Scene

Video-based vehicle target detection and tracking has broad application value,and has always been a research hotspot in the field of computer vision.However,in high-speed outfield scenes,unstable factors such as camera shake,sudden light changes,and severe weather increase the difficulty of vehicle target detection.At the same time,multi-scale,variable scale,mutual occlusion and similar targets of vehicle targets increase vehicle target detection and targets.The difficulty of matching.Therefore,it is of great theoretical and practical significance to study the detection and tracking of multi-scale vehicle targets based on video-based high-speed field scenes.SNIPER sampling strategy is a common method for multi-scale target detection,but there are problems of model training difficulty and slow detection speed.Therefore,this paper improves the method of SNIPER sampling strategy to obtain positive and negative chips,reducing the difficulty of model training,and proposes a A target detection method that cascades trajectory prediction,foreground detection and Faster R-CNN improves detection speed and accuracy.At the same time,in order to solve the problem of target similarity and occlusion during target matching,this paper also proposes a vehicle target matching method that combines motion features,apparent features,and trajectory features,which improves the accuracy of target matching.The main work and contributions of the paper are as follows:(1)For the problem of single image scale when obtaining positive chips,the paper proposes an adaptive image scale method based on mean shift clustering algorithm,which reduces the difficulty of model training;for the problem of acquiring additional RPN networks for obtaining negative chips,the paper The SNIPER sampling strategy is proposed to reuse the RPN network in Faster RCNN,which further reduces the difficulty of model training.(2)Aiming at the problem of the slow speed of the SNIPER sampling strategy in target detection,the paper obtains the approximate position and scale information of the vehicle target based on the Kalman filter-based trajectory prediction result and the Vi Be algorithm-based foreground detection result,and uses this to obtain The chips to be detected replace the image pyramid with smaller chips,which can speed up the detection speed and improve the detection accuracy.(3)Aiming at the problem of vehicle target matching,this paper proposes a matching method that combines motion features,apparent features,and trajectory features.It adopts the idea of cascade matching and designs different matching strategies according to different states of the trajectory,taking into account the matching accuracy and matching speed.When calculating apparent features,this paper designs and trains Le Net convolutional neural networks on the basis of twin networks to solve the problem of similar target matching;when calculating trajectory features,this paper uses the DTW algorithm to calculate the relationship between the target trajectory and the centroid-like trajectory The distance solves the problem of long-term occlusion.Based on the above improvements,a new multi-scale vehicle target tracking algorithm based on improved SNIPER sampling strategy for target detection and multi-feature target matching is formed.The monitoring video of the highway outside the field was collected and tested.The results show that compared with other target detection and target tracking algorithms,the method proposed in this paper has improved speed and accuracy,and has achieved ideal tracking results in practical applications.