Design And Realization Of Depth Detection System Based On Stereo Vision

Depth detection based on stereo vision has a wide range of applications in aerospace,assisted driving,virtual reality and other fields.Many scholars abroad have done much research,practice and breakthrough for all aspects of the depth detection system.The problem to be solved is still how to meet the core algorithm of three-dimensional matching's accuracy and real time at the same time.In this paper,a simple depth detection system was built which can be summarized as follows: binocular correction according to the calibration results,stereo matching to obtain the disparity map,3D points cloud reconstruction to obtain the depth and a rough estimate of the ground height.For real-time considerations,this paper selects semi-global stereo matching algorithm,and the latest stereo evaluation system(Version 3)of Middlebury University is used for carrying out multiple experiments on this algorithm.The main research results of this paper can be summarized as follows:1)In the acquisition of the initial disparity map,the feature extraction method combines the relative gray difference and Census transformation.Many experiments were designed for the weight coefficient of the Census transformation.It was found that the optimal selection range of the Census transformation weight coefficient is affected by the fixed window radius and dynamic programming's penalty cost.With the increase of the weight coefficient to a certain extent,the details information of the object,including contour changed dramatically and small objects distributed intensively,is detected better.2)In obtaining the final disparity map,this paper improves the method of calculating the supra-pixel disparity plane based on RANSAC.It is proposed that the tilt plane is established after that filling the partial voids of the initial disparity map in the way of controlled interpolation.After improvement,object edge mosaic phenomenon basically disappeared,then the contour becomes more natural and sharp.3)In measurement of the ground height,A height histogram of points cloud is established by the weighted statistics method of the different weight coefficient according to different depth ranges.The height of the ground is the abscissa of the first number peak correspondly from the minimum to the maximum.