Research On Multi-object Vehicle Detection And Tracking Technology In Traffic Scene

Multi-object vehicle tracking technology plays an important role in improving the utilization efficiency of existing traffic surveillance video data,timely and accurately grasping road dynamic traffic information and operation of regional road.This article takes highway surveillance video as the main research object and implements a multi-object vehicle detection and tracking system in a traffic surveillance scene.The research work of this article is as follows:Object detection: In this article,the vehicle detection in the monitoring scene is based on the YOLOv3 object detection algorithm.On the basis of analyzing the data imbalance problem caused by the too small Bus sample size in the existing data set and its impact,this article expands the Bus sample by data augmentation,and achieves m AP up to 89.68 %.Multi-object tracking: Vehicle trajectory can provide almost all dynamic information of moving vehicles in traffic scenes,and it is an important basic data for traffic analysis.In this article,multi-object vehicle tracking is studied as follows: 1)Kernelized correlation filter(KCF)is improved,and the method of solving the influence of object scale deformation and occlusion factors is designed to adapt to the long-term object tracking.2)A single camera multi-object tracking framework based on the similarity measure of intersection over union(Io U)and the trajectory state transfer is designed.According to the matching condition and track state between the detection nodes and the existing trajectories,the multi-object vehicle track is processed and analyzed at the same time.3)A cross camera vehicle trajectory association method based on Euclidean measure of trajectory nodes is proposed to solve the matching problem between the existing vehicle trajectory under the current camera and the new vehicle trajectory under the adjacent camera.It completes the long-term and long-distance continuous tracking and trajectory acquisition of multi-object vehicle.Trajectory smoothing and traffic parameters extraction: The multi-object vehicle detection and tracking algorithm has the characteristic of unsmooth in the original vehicle trajectory.This article proposes a vehicle trajectory smoothing method based on Savitzky-Golay(S-G)filtering to improve the jitter and unsmooth phenomenon On this basis,the traffic parameters are analyzed and counted.