| Stereo matching,as an important content of binocular vision,uses binocular cameras to obtain left and right images,calculates the left and right parallax of spatial object points,and then obtains the depth information of vehicle obstacles to achieve detection function.However,due to weak texture,noise,light intensity and other factors,the accuracy of parallax map is not high,which makes stereo matching has limitations.Therefore,how to improve the accuracy of parallax map and realize the obstacle detection around the vehicle has become an urgent problem to be solved.Based on this,aiming at the low detection accuracy of obstacles around vehicles in the field of intelligent driving,this paper improved the stereo matching algorithm.Firstly,the principle of binocular camera is described in detail,and the stereo matching algorithm is studied and improved.Finally,the improved stereo matching algorithm is used to restore the three-dimensional information of the obstacles around the vehicle.The specific research work and main achievements are as follows:(1)The basic theory of binocular stereo vision is introduced from the aspect of principle.Including binocular vision imaging principle and system composition,the transformation relationship between the four coordinate systems,comparing the camera calibration method and choose Zhang Zhengyou calibration method to obtain internal and external parameters,using Bouguet algorithm for stereo correction of the image,and stereo matching algorithm research points,specific steps,evaluation criteria,stereo matching to do a detailed system of research work.(2)In view of the traditional AD-Census algorithm in weak texture region matching accuracy is not high,support window setting is not ideal,this paper proposes an improved AD-Census stereo matching algorithm.In the calculation stage of matching cost,the weighted fusion of mean pixel instead of center pixel,RGB converted to HSV and gradient transform is used as matching cost,so as to improve the matching accuracy in this stage.In the cost aggregation stage,the improved cross crossing method is adopted to improve the matching accuracy of weak texture areas.In the stage of parallax calculation,the "winner-takes-all" algorithm is used to calculate the optimal parallax value after aggregation.In the parallax optimization stage,the invalid points are eliminated by using unique constraints,and then the background interpolation method is used to fill and optimize the four-direction scan line to reduce matching errors.Finally,the obtained parallax map is analyzed by qualitative and quantitative experiments on KITTI open data set,which shows that the matching accuracy of the improved AD-Census algorithm is improved.(3)Aiming at the problems such as poor effect of target area when the improved parallax map is used to detect obstacles in three-dimensional space,an obstacle detection method based on the combination of parallax map and V parallax map is proposed.Firstly,in order to eliminate the interference of road to obstacle detection,a road slant is obtained by using random sampling consistency(RANSAC)algorithm and Hough transform,and then the road is separated from the obstacle.Secondly,the direction and speed of moving obstacles are predicted,and the moving speed and direction are calculated by the change of object displacement in a very short time.Finally,simulation experiments are carried out to verify the reliability of this method in the process of vehicle obstacle detection. |
PDF Full Text Request