3D Vision-based Perception and Modelling techniques for Intelligent Ground Vehicles

Abstract

In this work the candidate proposes an innovative real-time stereo vision system for intelligent/autonomous ground vehicles able to provide a full and reliable 3D reconstruction of the terrain and the obstacles. The terrain has been computed using rational B-Splines surfaces performed by re-weighted iterative least square fitting and equalization. The cloud of 3D points, generated by the processing of the Disparity Space Image (DSI), is sampled into a 2.5D grid map; then grid points are iteratively fitted into rational B-Splines surfaces with different patterns of control points and degrees, depending on traversability consideration. The obtained surface also represents a segmentation of the initial 3D points into terrain inliers and outliers. As final contribution, a new obstacle detection approach is presented, combined with terrain estimation system, in order to model stationary and moving objects in the most challenging scenarios.