Real-time Obstacle Perception and Distance Estimation Based on Binocular Vision

To address the travel safety challenges faced by visually impaired individuals in complex environments, this study proposes a real-time obstacle perception and distance estimation method based on binocular vision. The system employs an improved YOLOv11n model to achieve high-precision obstacle detection (mAP@0.5: 82.83%) on the WOTR dataset, combined with a lightweight stereo matching network LightStereo for depth estimation. Through TensorRT FP16 quantization technology, the inference speed of object detection and stereo matching is improved by 3.2 times and 3.4 times respectively. The system is deployed on the NVIDIA Jetson Orin NX platform, capturing images in real-time through binocular cameras, obtaining obstacle categories and position distance information using object detection and stereo matching models, and outputting information through voice modules. The measured end-to-end inference latency is 46.76 ms, with absolute distance measurement error less than 0.25 meters within a 0-4 meter range, meeting the real-time and accuracy requirements for travel assistance for visually impaired individuals.