An effective obstacle detection system using deep learning advantages to aid blind and visually impaired navigation

Abstract

Blind and visually impaired people face different challenges when navigating indoors and outdoors. In this context, we suggest developing an obstacle detection system based on a modified YOLO v5 neural network architecture. The suggested system is capable of recognizing and locating a set of landmark indoor and outdoor objects that are extremely useful for Blind and Visually Impaired (BVI) navigation aids. Training and evaluation experiments were conducted using two datasets: the IODR dataset for indoor object detection and the MS COCO dataset for outdoor object detection. We used several optimization strategies, such as model width scaling, quantization, and channel pruning, to guarantee that the suggested work is implemented in embedded devices in a lightweight manner. The proposed system was successful in achieving results that were extremely competitive in terms of processing time as well as the precision of obstacle detection.