Design of a Blind Guidance System Based on RealSense and the Improved YOLOv5 Algorithm

To assist blind people in travelling and solve the problems of high hardware costs, insufficient portability of blind assistance devices, and vulnerability to environmental impacts, a modified YOLOv5 algorithm (YOLOv5-CM) based on ground plane segmentation and Euclidean clustering algorithms is proposed and applied to a blind guidance system based on RealSense D435. This proposed algorithm adds coordinate attention and uses MobileNetv3 as the backbone network for YOLOv5s to extract the main features. Compared with those of the YOLOv5 model, the mAP was improved by 0.7%, the model size was decreased by 79.1%, and the number of parameters was reduced by 80.3%. In this study, the YOLOv5-CM algorithm was applied to design a blind guidance system. When the system detects a traffic light, the RealSense D435 is used to obtain depth images to detect and measure the distance from the pedestrian traffic light. The system provides warnings based on the measured distance. When the system does not detect a traffic light, it utilizes the ground plane segmentation and Euclidean clustering algorithms to obtain the location and distance of obstacles. This information is conveyed to the blind through voice guidance, which aids navigation.