Visual depth estimation and 3D reconstruction using stereo vision for autonomous vehicles

Abstract

Stereo vision involves the estimation of disparity map by performing stereo matching between the left and right image pairs and reconstruction of 3D global points from the calculated depth points using the disparity. This dissertation explores five different algorithms starting from basic Block matching algorithm to some state of the art algorithms like Semi global matching with census transform combined with Slanted plane smoothing segmentation and it also includes the execution and analysis of these algorithm based on tested results in real time images. A new approach combining the selected positive points of two state of the art algorithms (ELAS+SPS) for stereo matching is proposed as a part of this dissertation and its performance is analyzed both qualitatively and quantitatively based on its disparity estimation by comparatively evaluating with both state of the art and traditional methods executed in this dissertation. 3D reconstruction of these stereo image pairs using the disparity map generated by all the discussed algorithms including the proposed algorithm has been executed and depending on the point cloud data generated from the disparity maps generated by all the discussed algorithms, mean depth error percentage is calculated comparing with the ground truth global points. 3D visualization using (PCL) Point Cloud Library is implemented for all the point cloud data generated by the different stereo matching algorithms. CGI and real time KITTI dataset images are used in the entire process of execution and testing of these algorithms. The main purpose of this dissertation is to estimate 3D global points with depth values with least processing time for guiding the autonomous vehicles.