Research On Visual Perception Technology For Vehicle Road Environment Based On Deep Learning Framework

Driving environment perception under complex conditions is a hot subject in current vehicle-assisted driving research,which is also one of the problems in the transportation field.Aiming at the information diversity of complex driving environment,this paper proposes a sensing method for multiple imaging target depth sparse optimization classification recognition in vehicle environment,and carries out researches on driving environment characterization based on different target perception.The paper aims at the target classification of images in the road environment,focusing on the pre-processing of vehicle vision images,target classification in vehicle road scenes,target boundary optimization,and visual synergistic perception of driving environment representation.The paper provides some theoretical support for intelligent assisted driving.In the aspect of perception in complex driving environment,firstly,the paper constructs the scheme framework of vehicle road image,and selects relevant devices.The images captured by camera are susceptible to noise,which makes the boundary between target and background is not obvious.This paper uses bilateral filtering method to remove noise from the image,and proposes an image enhancement algorithm for color image,which brighten the whole image and enhanced the contrast of edge detail.Secondly,it is difficult to use single feature to classify and characterize objects in an image,due to the complexity of vehicle-road and the variety of targets,a multi-source fusion and multi-scale pooling image classification algorithm is proposed.By adding gray image and canny operator transform input image in the network,the image features is enriched and multi-scale pooling is introduced to extract image details.The global information is obtained by feature fusion,which enhances the ability of the network to recognize vehicle environment targets and improves the accuracy of target recognition.Thirdly,in the case that the classified image has been mis-classified in target boundary and local small area,a method is proposed basing on the combination of super-pixel and conditional random field,which combines the super-pixel to re-classify the details and boundary of image target classification,then refines the image boundary using conditional random field algorithm to improve the image,which improve the overall performance indicators of classification.Fourth,in the aspect of vehicle environmentcharacterization,the occupancy grid map theory based on Bayesian reasoning is used to construct the occupancy grid map of the effective Lane area and updates it in real time.The Bayesian probability is used to calculate the status of each grid unit and to detect the dynamic obstacles in the lane.Finally,the experimental results show that the robustness and accuracy of vehicle environment target recognition are improved through the research,and the results have certain practicability.