Research On Vehicle Detection And Tracking Method Based On Convolutional Neural Network

Intelligent transportation is a major task in the strategic construction of China’s transportation power,and its application fields cover major national industries such as road transportation,railway transportation and aerospace.The number of motor vehicles in China was 39.3 million by the end of 2021.The rapidly increasing number of motor vehicles makes road traffic problems increasingly serious and hinders the development of intelligent transportation.Intelligent vehicle supervision system based on intelligent sensor equipment and intelligent algorithms has become the solution and is also a key part of the construction of intelligent transportation.Vehicle target detection and tracking technology is one of the key technologies of the intelligent vehicle supervision system.It collects road surveillance video by intelligent sensor equipment and analyzes the video image based on image processing technology,so as to guide road traffic operation automatically.How to design an intelligent algorithm for accurate detection and effective tracking of vehicles in road surveillance video is the main content of this thesis.At present,some researchers have explored vehicle target detection and tracking technology.However,the existing research work still has the following deficiencies:(1)the detection accuracy of existing vehicle target detection methods for small target vehicles on the road is low,which cannot meet the requirements of scenarios with high algorithm accuracy such as automatic driving.(2)The tracking accuracy of existing vehicle target tracking algorithms is low,and the problem of identity switching caused by vehicle target occlusion cannot be solved well.In view of the above problem,in this thesis the related theory of target vehicle detection and tracking technology was studied.We put forward a fusion attention mechanism of small target detection method of motor vehicles named YOLOv5-NAM and a real-time multi-target vehicle tracking method based on JDE algorithm named JDE-YN,improve the detection precision of the small target vehicle road and multi-target vehicle tracking accuracy,reduce the number of the vehicle target identity switches.The main works and contributions of this thesis are as follows:(1)In this thesis,the YOLOv5 object detection algorithm is extended,and a vehicle target detection model YOLOv5-NAM is proposed for vehicle detection of small road targets.Its loss function is optimized,and the network model is retrained on the UA-DETRAC vehicle target detection dataset.The m AP value of the new model is 1.6% higher than that of the original YOLOv5 s network model,and the detection accuracy of a single vehicle target is above 0.9 on average.Aiming at the problem of missing detection of small target vehicles in dense scenes,a SD-NMS non-maximum suppression post-processing method based on the penalty function idea and DIo U-NMS method was proposed,which improved the missing detection of small target vehicles.(2)In this thesis,JDE real-time multi-target tracking algorithm is extended,a real-time multitarget vehicle tracking method JDE-YN based on JDE algorithm is proposed,and the network model is retrained on UA-DETRAC vehicle target tracking dataset.The MOTA value of the new model is improved by 0.9% compared with the original JDE algorithm.Aiming at the problem of identity switching of vehicle targets in the process of vehicle target tracking,a cascade matching algorithm based on direction correction was proposed,which reduced the number of identity switching of vehicle targets by 15%.(3)In this thesis,we apply the improved algorithm proposed above to the intelligent park road vehicle supervision system,design and implement parking space management service based on vehicle target detection and vehicle flow statistics service based on vehicle target tracking.Experimental results show that the proposed method can effectively detect small target vehicles and track multi-target vehicles in real time and efficiently,which can promote in-depth research in the field of vehicle target detection and tracking.