Embedded implementation of an obstacle detection system for blind and visually impaired persons' assistance navigation

Blind and Visually Impaired (BVI) persons encounter safety problems during their navigation. Therefore, assisting BVI must be addressed. Obstacle detection and avoidance in real scenes present very challenging tasks. To handle this challenge, we suggested developing a new obstacle detection system based on an enhanced YOLO v5 neural network. The improved network architecture increased both the network's speed and the detection accuracy. This was achieved by integrating the DenseNet into the YOLO v5 backbone, which impacted the reuse of features and data transfer with additional modifications. Aiming to ensure an embedded implementation of the proposed work on a ZCU 102 board, we applied two compression techniques: channel pruning and quantization. The performance of the suggested system in terms of detection and processing speed showed very encouraging results. In fact, it achieves a detection accuracy of 83.42% and a detection speed of 43 Frame Per Second (FPS).