Energy Minimization Approach for Negative Obstacle Region Detection

Obstacle detection is an important area of research for automatic robot navigation and obstacle avoidance. However, the main focus has been on detecting positive obstacles such as vehicles, pedestrians and anything above the ground plane. Equally important task is detecting negative obstacles such as holes, ditches etc. that lies below the ground surface, which has not been studied extensively. We propose to detect such holes and potholes as an energy minimization problem. We use stereo information combined with saliency to initialize the energy function and use colour information to optimize the result. The optimized result is then subjected to hysteresis thresholding to arrive at the final negative obstacle region. Experimental results prove that the introduced method can successfully detect negative obstacles and run in real time (in the range 14-20 Hz) when using GPU for stereo matching. The proposed method was tested under three camera configurations with corridor, concrete-road and tar-road scenes. Proposed method perform well compared to the recent works.