Study On All-weather Road Environment Perception Technology Based On Vehicle Vision

Road environment perception systems have been widely studied by researchers as an important part of automotive auxiliary systems.The road environment perception system established in this thesis mainly includes lane line detection and moving object recognition.Most researchers use convolution neural network algorithms to achieve two tasks.However,these algorithms have some limitations in real-time performance,robustness and memory.In this thesis,the practical application of road environment perception system under complex weather in traffic scene is studied,and an all-weather road environment perception system is proposed.The main work of this thesis is as follows:(1)Based on cascaded multi-task network,a network Cascading D-SNet(Cascading Detection and Segmentation Network),which can detect object and identify lane simultaneously is established.The accuracy of lane recognition is increased by 3.5%,and the running time is 73.72 Ms less than the sum of the two tasks.(2)Using the idea of parallel multi-task network,,the multi-weather network structure MWNet(Multi-Weather Net)is established.The universality and practicability of the network structure are proved by two tasks of object detection and lane segmentation.The accuracy of lane line recognition and object detection under bad weather can be improved by 3.6%and 3.3%respectively.(3)To meet the requirements of robustness,real-time performance and hardware memory capability of the road environment perception system,an all-weather road environment awareness system is established.The system realizes five tasks:real-time classification of image weather,object detection and lane line recognition under bad weather,object detection and lane line recognition under good weather conditions by sharing a basic network.On the basis of ensuring the detection accuracy and running time,the memory required by the algorithm can be further reduced,and the practical application ability of the algorithm can be improved.(4)At present there is no database of road route and moving objects in road traffic under various weather conditions.In order to verify the performance of the network established in this thesis,data collected from 7 cities across the country are used to establish a database of more than 60000 samples.This data set has great application value and reference significance for studying the object detection and lane line recognition under various weather conditions in China.In summary,on the basis of computer vision technology,we studies the optimization of object detection and lane recognition algorithm based on convolution neural network under various weather conditions in complex traffic environment.It provides a new idea for the practical application of the algorithm.