Traffic object detection for fixed-camera expressway surveillance under degraded imaging conditions

Traffic object detection in fixed-camera expressway surveillance is challenging because wide-view monitoring scenes contain mixed object scales, distant small targets, weak visual evidence, and degraded imaging factors such as low resolution, blur, weak contrast, glare, and background interference. This study aims to improve multi-class traffic object detection under these practical surveillance conditions while maintaining deployment-oriented real-time inference efficiency. A surveillance-oriented detector is developed on the basis of YOLOv5s. The method introduces two task-driven mechanisms. First, a fine-scale P2 prediction head is added to provide an additional high-resolution response path for small and distant traffic objects. Second, context-aware reassembly of features (CARAFE)-based feature reassembly is incorporated into the neck to improve cross-scale feature fusion for visually weak objects. A self-built expressway surveillance dataset was constructed from roadside videos collected on the Chengdu-Chongqing Expressway. The dataset contains 53582 images, and low-resolution and blurry samples were intentionally retained to reflect realistic monitoring conditions. Experimental results showed that the proposed detector achieved 66.71% mean average precision (mAP), 60.23 frame·s^-1, and 50.32% mAP on the small-object subset under the reported evaluation setting. Compared with baseline YOLOv5s, the proposed detector increased overall mAP from 61.89% to 66.71% while maintaining practical inference speed. Controlled module-level analysis on a representative subset further showed that the P2-only setting increased small AP@0.5 from 12.83% to 17.15%, indicating its direct contribution to fine-scale target sensitivity. These results suggest that the proposed detector improves surveillance-oriented traffic object detection in the constructed fixed-camera expressway setting. The evidence is mainly grounded in the self-built application dataset, but broader cross-road and cross-scene validation remains a direction for future work.