Research On Structured Road Perception Method Based On Binocular Vision

Driving conditions in urban roads have become more and more complicated,causing a large number of traffic accidents.To help solve this problem,researches on Advanced Driver Assistance Systems and Unmaned Intelligent Vehicle Systems become popular,which also promote the development of Pattern Recognition,Computer Vision,Intelligent Control and etc.A tough and critical problem remains in such systems that how to use the technology of computer vision to achieve an accurate environment sensing in complicated urban areas.In this paper,an improved vanishing-point detection method,a road segmentation method and an obstacle detection method using stereo vision system and machine learning methods are proposed.The proposed methods improve the performance of the environment sensing systems.Below are our three main contributions:(1)An improved vanishing point detection method is proposed.By analyzing the U-disparity map,an improved V-disparity map is acquired to improve the accuracy of the detection of the vanishing line.And then,a vanishing-point constrained Dijkstra road method(Vp-ShortestPath)to segment the road region is proposed.The Vp-ShortestPath method is a nonparametric model,which can detect precise road borders without assuming parametric equation.In some cases,the Vp-ShortestPath method is superior to many deep-learning-based methods.(2)In this paper,a Disparity-Aware-Loss fully convolutional neural network(DAL-FCNN)is proposed for road segmentation task.The DAL-FCNN can improve the troditional FCNN model.Then,a framework of conditional random fields is prospoed to fuse the DAL-FCNN and Lidar to enhance the road segmentation result.(3)A convex-obstacle detection method is proposed in U-disparity map,which is based on the Vp-ShortestPath method.This convex-obstacle detetcion method is efficient to detect the vehicles in front of the camera.This method can provide effective decision for environment sensing system of the vehicles.