Fast optical flow estimation and its application to real-time obstacle avoidance

Abstract

This paper presents a novel fast optical flow estimation algorithm and its application to real-lime obstacle avoidance of a guide-dog robot. The function of the laboratory-developed robot is to help blind or visually impaired pedestrians to move safely among obstacles. The proposed algorithm features a combination of the conventional correlation-based principle and the differential-based method for optical flow estimation. Employing image intensity gradients as features for pattern matching, we set up a brightness constraint to configure the search area. The merit of this scheme is that the computation load can be greatly reduced and in the mean time the possibility of estimation error is decreased. The vision system is installed on-board the robot to provide depth information of the immediate environment. The depth data are transformed to a safety distribution histogram and used for real-time obstacle avoidance. Experimental results demonstrate that the proposed method is effective for a guidance robot in a dynamic environment.