Research On Preceding Vehicle Detection Based On Deep Learning

With the development of urbanization and the popularization of vehicles,as well as the frequent traffic accidents,the safety of driving has become a global concern.In this context,the development of automotive intelligence is particularly important.Advanced Driver Assistance System(ADAS),as an important component of Intelligent Vehicle System(IVS),can effectively improve the safety performance of vehicles and improve the intelligence of vehicles.As the main function of ADAS,real-time detection of in front of vehicle can effectively "predict" the hidden collision dangers and promptly give the driver a timely warning,which is one of the key technologies to achieve vehicle anti-collision,effectively reducing the proportion of casualties in the road traffic accidents.Firstly,this paper analyzes the obvious characteristics of the shadow of the vehicle under the road,and in order to reduce the impact of the illumination change on the robustness of the algorithm,a shadow segmentation algorithm based on monocular vision is proposed.Then,we divide the image environment into three categories: strong light,normal light,and weak light.To reduce the interference factors in the non-road environment region,a road region detection method based on edge enhancement is proposed.Based on the road detection result,an adaptive shadow segmentation algorithm is proposed for daytime road environment.Deep learning is a feature learning approach that transforms raw data into a higher-level and more abstract expression through a simple,non-linear model.In view of the the characters of sparse connection,weight sharing and down-sampling of the convolution neural network method,this paper discusses the training method of the parameters and the structor designing.The algorithm is applied to express the characteristics of vehicles.Based on the shadow segmentation results,a deep learning model is used to validate and remove the limited number of candidate regions of vehicle to realize the accurate recognition and location of the vehicle.Finally,this paper tests the visual images of different expressway or urban road scenes to verify the general validity and adaptability of the detection algorithm the paper proposed.Meanwhile,according to the three conditions mentioned in the long-distance road test,the missing rate and false detection rate for each case are calculated respectively.The experimental results show that the algorithm proposed in this paper has good adaptability to complex environment,and can detect the moving vehicles ahead accurately and effectively.In addition,this method is compared with the traditional monocular vision baesd vehicle detection algorithm to further verify the advantages of this algorithm.