Two-Stage Obstacle Detection Based on Stereo Vision in Unstructured Environment

Abstract

In an unstructured environment, there are many challenges for obstacle detection. This paper presents an improved method to detect obstacles based on stereo vision in unstructured environments based on salient obstacle extraction. This method can achieve same or higher level of accuracy of obstacle detection compared to the existing salient obstacle detection with significant reduction of computation time. This method consists of two stages. In the first stage, it extracts the salient obstacles which stand out from the background in the stereo images using a fast salient obstacle detection method. In the second stage, it refines the detection of small obstacles by computing the geometric relationships among 3D points using an improved space-variant resolution (SVR) with the continuity and the height constraints. The experiment results show that this improved method can reduce computation time and improve detection accuracy.