Intelligent Obstacle Avoidance System Based on Binocular Vision and Ultrasonic Fusion for Visually Impaired People

Abstract

Aiming at the shortcomings of current visual impairment assistance devices in transparent obstacle recognition, low-lying object detection, and dynamic target perception, this paper designs and implements an obstacle avoidance system fusing binocular vision and ultrasonic information. The system adopts a master-slave embedded architecture with Raspberry Pi 4B and STM32 as the core. An improved semi-global matching algorithm is introduced at the vision end to support medium and long-distance semantic perception. Meanwhile, combined with a five-direction ultrasonic array, high‑precision near-field detection within the range of 0.1–8.0 m is realized. Multi-source perception data are collaboratively fused in time and space through extended Kalman filter and occupancy grid map, and low-interference prompt information is output via bone conduction voice and vibration cane. Experimental results show that in complex indoor and outdoor environments, the average recognition success rate of the system for various types of obstacles reaches 98.5%, and the early warning delay is about 210 ms. The overall performance is superior to single-sensor schemes. The system does not rely on network and has controllable cost, providing a feasible implementation for universal accessible intelligent assistance equipment.