Research And Implementation Of Vehicle Tracking Based On FCOS

In today's society,with the rapid development of science and technology and the improvement of people's living standards,the number of vehicles keep increasing,which leads to intense pressure on transportation.Therefore,the Intelligent Transportation System(ITS)has come into being.ITS uses computer vision,artificial intelligence and other technologies,helps people to supervise and manage traffic,identify road congestion,traffic accidents and other issues in a timely manner and take measures to reduce the investment of human,material and financial resources by utilizing the high-performance processing capabilities of computers,thereby making urban traffic more safe and smooth.Recognition and tracking of vehicles in road surveillance videos is one of the most important contents in ITS,and has attracted widespread attention from scholars at home and abroad.This thesis studied the vehicle tracking problem in a single camera scenario,and the overall algrorithm framework included two parts,vehicle detection and vehicle tracking.In terms of vehicle detection,this thesis first compared traditional vehicle detection algorithms with deep learning based detection algorithms,then propose to use fully convolutional one-stage object detection(FCOS)algorithm for vehicle detection,and maked comparisons with the other widely used vehicle detection algorithms from the experimental points of view.Finally,the experimental results verify the applicability of FCOS algorithm in vehicle detection.As for vehicle tracking,it mainly included two parts: feature extraction and vehicle association.In order to more accurately compare vehicle similarity,this thesis analyzed a variety of feature extraction methods and combined the characteristics that vehicle images in domestic road monitoring videos are prone to feature mutations,then proposed to combine Face Net deep learning features and edge features as vehicle appearance features.On the problem of vehicle association,due to the complexity of the actual road scene,the mutual occlusion between vehicles makes it difficult to track the vehicle.Thevehicle tracking method cannot handle the occlusion problem well only by comparing the similarity of vehicles in adjacent frames.This thesis used the intersection over union and feature similarity comparison results between vehicles in the previous and subsequent frames to initially generate tracklets of the vehicles on a fixed timewindow,and combined the tracklet graph model and the Tracklet Net trajectory similarity measurement network for vehicle association,thereby effectively reducing the impact of occlusion.The effectiveness of the vehicle tracking algorithm proposed in this thesis was also verified by experiments.Finally,this thesis designed a vehicle tracking system with a vehicle detection module and a vehicle tracking module as the core.The proposed vehicle tracking algorithms were used in the system building process to ensure the accuracy and improve the operation speed as much as possible.Finally,the accuracy,validity and practicability of the vehicle tracking system designed in this thesis were verified by testing.