Research on Tunnel Pedestrian Detection Algorithm Based on Image Enhancement and Threshold Segmentation

ABSTRACT As an important infrastructure that can shorten the travel distance, improve the transportation capacity and reduce traffic accidents, tunnels are faced with the potential safety hazard of pedestrians crossing without authorisation, which often leads to major traffic accidents. Consequently, a low‐light image enhancement and denoising algorithm tailored for tunnel environments using computer vision and deep learning technology is proposed, which refines the representation of characteristics in partially dark areas by examining relevant parameters and spatiotemporal features of raw tunnel scene video images. Building on this foundation, a pedestrian detection method based on connected domain dynamic threshold segmentation is designed to reduce the computational resources for identifying pedestrian positions by adopting foreground segmentation based on a fixed background. Furthermore, a lightweight classification neural network is integrated to classify candidate regions for confirming pedestrian presence. Experimental results demonstrate the algorithm's superiority in terms of detection efficiency and accuracy. Compared with the baseline models, the enhanced results of DA‐Zero‐DCE proposed in this paper are improved by about 10 dB and SSIM by about 0.1. For the proposed DAZ‐CCDS detection model, the test accuracy is improved by about 10% when the confidence threshold is 0.75 compared with the mainstream target detection model.