Single-camera Stereo Vision And Its Application Research In Road Scenes

Over the past years,with the increase in car penetration rate,the traffic environment is getting more and more complicated.As a result,researches on Unmanned Intelligent Vehicle Systems have become the focus of Artificial Intelligence.Realizing the 3D sensing of the driving environment with the help of Computer Vision technology and Pattern Recognition technology is crucial to these systems.Camera,lidar and radar are common sensors in these tasks.Comparing with other kind of sensors,cameras are more cost-efficient and spacesaving,and they can obtain abundant color information.How to recover the 3D position of the landmarks through images is the key problem to schemes based on vision.Stereo cameras are usually expensive,and they need extra technologies to synchronize two different cameras.How to build a stereo vision system with one camera and recover the depth information of landmarks accurately has interested researchers.This paper focuses on single camera stereo vision system and its application on road scenes.Below are our main contributions:(1)This article proposes a single-camera stereo apparatus.This apparatus combines one camera and four mirrors to achieve the function of a stereo camera.Through the reflection of lights,this apparatus can capture a pair of images in stereo vision in one frame.Our device is cheap and compact,and it can be attached to a single-camera to build a virtual stereo system.It has a suitable field of view for road scenes.Our experimental results demonstrate that this system can be applied to the road environment sensing tasks for intelligent vehicles,and it has a pleasing performance on vanishing point detection and road detection;(2)This article proposes a novel algorithm for the self-calibration for the camera-toground parameters,which optimizes the camera-to-ground parameters with the observation information of landmark points and the motion information from a set of frames.The camera-to-ground parameters represents the relative position between the road surface and the camera,and the stereo vision can be achieved by a single camera through the parameters.The accuracy of the parameters determines the accuracy of 3D position of landmarks.Our experimental results demonstrate that our algorithm can obtain accurate camera-to-ground parameters,and 3D points can be estimated accurately with these parameters.