POP-YOLOv8: an object detection framework for partially occluded pedestrians in nighttime traffic environments

Nighttime traffic environments are filled with potential hazards, among which pedestrians are not only among the most frequently appearing targets but also some of the most easily overlooked moving obstacles by drivers in dark environments. To enhance driving safety under such conditions, we propose POP-YOLOv8, a customized object detection model designed specifically for detecting partially occluded pedestrian detection in nighttime traffic scenes. To improve both detection accuracy and real-time performance of YOLOv8n under low-light and occlusion-prone conditions, four key enhancements are introduced: a Feature Enhancement Module (FEM) is embedded in the backbone to better capture the discriminative features of occluded pedestrians. An POPAM attention mechanism is incorporated into the neck to emphasize salient regions in complex traffic scenes. GhostCSB and C2fGhost modules are integrated to reduce parameter redundancy and enhance model efficiency. In the inference phase, the most critical SCI module operates in FP32 precision, while other layers use FP16 to balance accuracy and speed. Additionally, ECA and residual structures are incorporated to preserve image detail during inference. Comparative experiments and visual analyses demonstrate that POP-YOLOv8 outperforms YOLOv10 in detecting partially occluded pedestrians under nighttime traffic conditions, providing superior robustness and computational efficiency.