关键词: 户外, 盲人导航, 实例分割, RGB-D, 距离检测, 离群点剔除, BlendMask, 深度学习

Abstract:

In view of the problems such as unreasonable design, incomplete construction of barrier-free facilities, uneven guide AIDS, single functions and inability to meet users' outdoor travel needs, this study proposes a visual perception and detection method based on BlendMask algorithm and RGB-D dynamic obstacle targets for outdoor guide for the blind. The proposed method uses RGB-D data collected by Azure Kinect DK, compares six sets of instance segmentation algorithms, selects the best BlendMask instance segmentation model, and adds a depth data processing branch to obtain the depth information of a single instance, that is the distance from the target to the camera. To improve distance detection accuracy, outlier rejection operations are used in 3D, in the form of point clouds, to remove residual background information from the segmentation mask, where a statistically based outlier rejection method can more accurately detect the distance from the obstacle to the camera.Experimental results show that the proposed method is not only better at detecting dynamic outdoor targets, but also detects the distance from the target to the camera and improves the accuracy of distance detection.

Key words: outdoor, navigation for blind people, instance segmentation, RGB-D, distance detection, outlier rejection, BlendMask, deep learning