Fast obstacle detection using U-disparity maps with stereo vision

This paper introduces a computationally efficient, approach for obstacle detection in driving assistance applications, based on stereovision. The proposed approach involves three different steps aiming an increased quality of the results. The first step relies on the basics: obstacles correspond to peak regions in the u-disparity map. By applying the model of the stereo system, the peaks are detected with an adaptive threshold. The adaptive thresholding will calculate the accumulation of points required for an obstacle based on its distance (disparity) and will be related to the characteristics of the stereovision system. The second and third steps consist of refining the result of the previous, vertically respectively horizontally. This is necessary in order to fill out unmarked pixel regions which are classified as belonging to obstacles. The second step iterates vertically and propagates the obstacle label to neighbor pixels. The third step improves obstacle regions horizontally, with points that do not belong to the road. The solution is fast and reliable, on various scenarios, as every step is an improvement of the standard U-disparity approach for obstacle detection.