Obstacle Detection for Mobile Robots, Using Dense Stereo Reconstruction

Abstract

Sparse (edge based) stereo reconstruction has been used for many years, being less computational expensive. The dense (all pixels in image) stereo reconstruction, real-time computable nowadays, brings 3D reconstructed points even for less textured image regions, and has a lower percentage of wrong reconstructed points. This article presents novel algorithms for obstacle detection, using dense stereo reconstruction. They analyse, on the top view of the scene, the local density and vicinity of the 3D points, and determine the occupied areas which are then fragmented into primitive obstacles based on a concavity-free criterion. The obstacles are modelled as cuboids, and their orientation is determined in order to get a very good modelling of the obstacles in the scene ahead and, consequently, to minimize the free space which is encompassed by the cuboids. The main abilities of the approach are: generic obstacle detection, determination of obstacles' orientation, confident fitting of the cuboidal model.