An Obstacle Detection Method for Visually Impaired People Based on Semantic Segmentation

Using low-cost visual sensors to assist indoor and outdoor navigation is an important method to solve the problem of visually impaired people living and going out. To this end, we proposed an obstacle-detection method for visually impaired people based on semantic segmentation. We use the semantic segmentation method to determine which targets in the camera view field need to be noticed and use the related information to establish a real-time local map. At the same time, we propose a method to fuse semantic information with local point clouds, achieving obstacle detection based on probability fusion. Finally, the distance between the interested target and the camera will be returned and sent to the user. The proposed method can achieve visual navigation with more than ten frames per second (fps), lower than 0.3 m detection accuracy, and smaller than 4 MB generated model, which is also compatible with multiple cameras and control terminals.