Research On Road Vehicle Detection Method Based On Deep Learning

With the fast development of social economy,the rapid popularization of urbanization and automobile,traffic congestion,frequent traffic accidents,increasing urban traffic pressure and other issues have become common problems faced by all countries in the world.Road vehicle detection technology is the basis for traffic flow statistics,road congestion analysis and other tasks to obtain basic information such as vehicle number,location,vehicle type,etc.,which provides data support for vehicle behavior interpretation and traffic congestion analysis.Therefore,vehicle detection is a basic and key work of intelligent traffic information acquisition,which has great practical significance and practical application value.The traditional road vehicle detection technology mainly adopts the method of selecting candidate region based on sliding window and extracting features by hand design,which is easy to cause the problem of low robustness and affect the detection performance and efficiency.With the development of deep learning technology,the target detection technology based on deep neural network has made a significant breakthrough in detection performance and efficiency by designing different network models to automatically learn more effective vehicle features.However,in the real monitoring scene,the road vehicles in different light conditions and shooting angles show great differences in appearance and scale changes,which lead to the existing road vehicle detection technology to be faced with great challenges.Therefore,starting from the characteristics of real road scene vehicles,this thesis explores the road vehicle detection method based on deep learning,and proposes an improved road vehicle detection method.The main research works are as follows:(1)Through the investigation of domestic and foreign literatures,this thesis focuses on the analysis of the most representative convolution neural network structure and training methods in deep learning,and summarizes the current main road vehicle detection methods in theory and model respectively,carries out the detection experiment and effect analysis of these road vehicle detection methods,which lay the relevant foundation for the follow-up research.(2)Road vehicles show large scale differences and appearance changes in different shooting angles,which makes the existing vehicle detection methods still to face challenges in the expression of the target area.In this thesis,it is designed that a more robust road vehicle detection method based on multi-level feature fusion network.This method is based on the regional convolution neural network.By using the multi-scale vehicle generator module to simulate the change of vehicle appearance scale in real situation,it enriches the depth characteristics of network learning.Combined with the prior anchor point generation method and multi-scale network structure,it improves the detection effect when the vehicle appearance is multi-scale difference.(3)In the real traffic environment,the distinguishability between the vehicle and the background is reduced and the appearance of the vehicle is blurred due to the change of illumination.These factors make the existing vehicle detection methods difficult in feature extraction.In order to solve this problem,this thesis proposes a road vehicle detection method based on visual perception network.Using the visual perception mechanism to learn the vehicle characteristic under various light conditions,combining with the one-stage detection network,while ensuring the detection efficiency of the whole detection method,it enriches the semantic information expression of the network,improves the effect of vehicle detection under different light conditions.(4)Aiming at all kinds of scale changes of vehicles in the road scene,most of the target detection methods adopt the way of setting anchor window empirically,which ignore the scale changes of vehicle diversity in the real scene and affect the detection effect and efficiency.Therefore,using the semantic information and spatial location information of the target,this thesis proposes a road vehicle detection method based on the self-adaptive anchor generation network.Through the self-adaptive anchor generation network,we can learn deeper features,and generate anchor windows that conform to different vehicle positions and shapes without manual intervention,so as to provide more accurate suggestion areas for detection,which can improve the detection effect of vehicle under multi-scale change.(5)The proposed methods and network models are trained and tested by using the open data sets.The experimental results show that the proposed methods are feasible,and the detection effects are improved compared with the representative detection method.In summary,the improved vehicle detection methods based on multi-level feature fusion network,visual perception network and adaptive anchor generation network proposed in this thesis have robustness to vehicle appearance scale shape,adaptability to environment and self-adaptability in multi-scale situation.