Collision detection method using image segmentation for the visually impaired

Abstract

To assist the visually impaired (VI), a variety of collision detection methods using monocular vision have been developed. Most conventional collision detection methods for the VI utilize feature points and their corresponding motion vectors. However, when the VI subject approaches a non-textured object/obstacle, such as a door or wall, the conventional methods often fail to detect the collision on account of insufficient feature points and inaccurate motion vectors. To address this problem, this paper presents a collision detection method using image segmentation. In the proposed method, the input frame is over-segmented into superpixels by using the superpixel lattices algorithm. The segmentation result is then obtained by applying a graph-based region merging algorithm to the superpixels. Finally, the collision is detected using the geometric relationship between the size variation of the image segment and the distance variation from the camera to that segment in a real-world environment. Experimental results demonstrate that the proposed method handles a variety of scenarios, including a non-textured object, while outperforming conventional methods in terms of accuracy.