Integrated u-net and SimAM-YOLO v5s for sidewalk and traffic light recognition

Visually impaired individuals face significant challenges in accessing affordable assistive technologies for environmental perception. This study proposes a computer vision system that integrates two deep learning architectures: an improved YOLO v5s model incorporating the SimAM attention mechanism to enhance traffic light recognition accuracy and a U-Net- based network for pavement segmentation. The experimental results demonstrate that the improved YOLO v5s achieves 93.5% precision (2.7% higher than baseline) with mAP50 of 92% and has better performance than the original. U-Net can solve the detection problems successfully. This approach offers a potential solution to overcome the high cost of assistive devices.