Obstacle detection using U-disparity on quadratic road surfaces

This paper addresses the problem of detecting obstacles that protruding from the road. Traditionally, the road surface has been considered flat, and camera orientation is fixed. However, both assumptions are not strictly true in urban scenarios. The proposed algorithm employs a time-of-flight (ToF) camera. It allows dynamic pitch/roll angles, height variations and represents the ground as a quadratic surface. The range information given by the camera is represented in both Euclidean and disparity domains, so that their domain characteristics support each other to achieve accurate and efficient detection results. Gradient filtering of the disparity image presents Euclidean planner patches, with which outliers can be minimised during road fittings. Obstacle points are subsequently detected by the connect component labelling algorithm. Experimental results show that the proposed method can effectively segment and detect multiple obstacles and presents their bounding boxes in complex scenarios.