Research On The Driving Environment Perception Method Based On Visual Cooperative Vehicle-Infrastructure System

Driving environment perception is the most core content in the Advanced Driving Assistant Systems(ADAS).The driving environment information is usually collected through on-board sensing devices.Meanwhile,intelligent roadside device is an important manner of environment information perception in Cooperative Vehicle Infrastructure System(CVIS).It can exchange information with intelligent on-board devices via Vehicle-to-Infrastructure(V2I)communication,so that the CVIS can get more comprehensive and rich driving environment information.However,as driving environment information is multifarious and variety,how to collect,process and fusing these information based on actual demand and achieve driving environment perception,is a very important research topic in ADAS.This paper propose an Visual Cooperative Vehicle-Infrastructure-based Driving Environment Perception System.On this basis,focusing on the system requirements on driving environment information perception,researching on lane marking and vehicle information sensing method,and driving environment characterization method based on the above-mentioned information.The main research findings include the following aspects:(1)Designing a structure of the Visual Cooperative Vehicle-Infrastructure-based Driving Environment Perception System.Based on in-depth analysis of the driving environment information perception and interaction requirements in ADAS,combining with the basic structure of CVIS,we propose the structure of the Visual Cooperative Vehicle-Infrastructure-based Driving Environment Perception System.In this system,roadside sensors and on-board sensors are mainly sensing devices.V2 I communication technology is used to transfer and exchange information.Intelligent roadside device and intelligent vehicle terminal are the core carrier of driving environment perception and interaction.Vehicle's self-state information and environment information such as lane marking and obstacle vehicles are perceived in this system.(2)Research on structured road lane marking identification method in on-board video.Based on traditional model-based lane marking detection and tracking method,we furtherconsider that lane marking type is important to measure if a lane change behavior is legal.So we implement a lane marking detection,classification and tracking method based on non-uniformed B-spline(NUBS)curve model matching.After obtaining the control points based on edge information of lane marking,we firstly formulate an strategy to classify lane marking into dashed and solid.Then we implement lane marking detection and tracking by NUBS curve restructure and curve estimation.Experiment results show that this method can detect,classify and track lane marking efficiently,and is robust to the case of partial edge information loss of lane marking.(3)Research on vehicle detection and tracking method in road-side video.Focusing on vehicle edge feature is easily interfered by background edge and noise,we propose an non-sampling Difference of Gaussian Pyramid multi-scale edge fusion-based vehicle detection method.After non-sampling DoG Pyramid decomposing,edge detection and edge fusion,the obtained multi-scale vehicle edge map inherit the advantages of large and small scale image edge,while eliminate the effect on edge detection caused by up and down sampling in DoG Pyramid decomposition.Executing morphological processing and connectivity analysis on the multi-scale vehicle edge map to detect vehicle.Experiment results show that the method can work effectively under different weather conditions.Traditionally,vehicle detection and vehicle tracking are achieved by different way respectively,and usually result in highly algorithm complexity.Thus,we propose an SIFT feature matching-based vehicle detection and tracking method.Two adjacent frame images align after SIFT feature matching and differ with each other to obtain a difference map.We search the region with a higher sum of absolute difference to detect vehicles.Furthermore,for vehicle tracking,we construct a vehicle tracking sample set,use the result of vehicle detection as a tracking sample to match SIFT feature with new frame.Meanwhile,a tracking sample set updating strategy is proposed to identify and process three special cases,including vehicle entering into,exiting from detection field,and stopping.(4)Research on driving environment representation method based on visual cooperative vehicle-infrastructure perception information.Firstly,we achieve an Occupancy grid-baseddriving environment representation method.In this method,we improve the point-to-point data mapping relationship in traditional Occupancy grid.We further consider the effect of vehicle's size on driving environment representation,and use an Gauss distribution to map a vehicle coordination point into grid.Bayesian Probability Theory is used to fuse road-side vehicle position information and on-board lane marking information,and compute probability of two state “free” and “occupied” for each cell.Experiment results show that the improved method can obtain more accurate driving environment representation.However,the computation cost of Occupancy grid is very high,and Bayesian Probability Theory cannot deal with uncertain.Thus,we further propose an dynamic credibility map-based driving environment representation method.We construct the map based on vehicle size and lane marking position.Then we use Dempster-Shafer theory to fuse GPS vehicle position information,vehicle detection and tracking information,and lane marking type information.Finally,we compute belief mass of three states “free”,“occupied” and “dangerous”.Compare with the former method,this method can more accurately represent driving environment.In order to applying validate the accuracy of the driving environment representation result and practicability of the driving environment representation method,we further achieve Bayesian network-based lane changing assistance decision and rules fusion-based lane changing assistance based on above-obtained representation result.The former can output three decision “No lane change”,“Lane change to left” and “Lane change to right” through Bayesian network computation.The decision with the highest expected utility is the optimal lane changing decision.The latter through the computation of Spatial Cost,Time to Collision,Required acceleration,to output more detailed lane changing decision with acceleration information.Experiment results show that output decision can guarantee driving safety while play to driver's lane change intention.It also validates the efficiency of the proposed driving environment perception method in this paper.