YOLO-Improved Lightweight Hybrid Models for Real-Time Blind Path Detection

Visually impaired individuals face significant challenges in outdoor navigation, including dynamic obstacles, complex terrains, and real-time requirements, necessitating the development of efficient assistive technologies. While traditional tools like white canes and guide dogs exhibit limited functionality in dynamic environments, existing deep learning models—despite their excellence in object detection and path segmentation—suffer from high computational complexity and excessive resource consumption, rendering them unsuitable for mobile device deployment. This study proposes a YOLO-MobileNet hybrid architecture to deliver a lightweight, real-time solution for blind path segmentation and obstacle detection. The model synergistically integrates detection capabilities of YOLO and efficiency-optimized design of MobileNet, achieving high performance on resource-constrained mobile platforms. We developed and optimized the YOLO11m-MobileNetv4 framework, trained and validated on a custom-built dataset for blind path segmentation and obstacle detection. Comparative experiments were conducted across multiple environments against YOLOv8, YOLO11, and their lightweight variants. Key evaluation metrics include mAP50, mAP50-95, FPS, and GFLOPs. Experimental results show that YOLO11m-MobileNetv4 achieved an mAP50 accuracy of 0.7323 in the obstacle detection task. Its FPS on the GPU reached 148.95, and the FPS on the CPU was 97.30% higher than the baseline model. At the same time, the number of parameters is significantly lower than that of the baseline YOLO model, showing excellent real-time and lightweight characteristics. This work presents a lightweight deep learning framework that enhances the practicality and scalability of visual impairment navigation systems. The proposed model not only advances outdoor safety and independence for the visually impaired but also exhibits broad application potential in mobile edge computing.