ENHANCED YOLO FOR REAL-TIME MULTI-SCALE TRAFFIC DETECTION UNDER HAZE CONDITIONS

In road traffic safety systems, computer vision-based object detection faces significant challenges in hazy weather, including large variations in scale, high background noise, and complex viewing angles. To tackle these problems, we propose an improved YOLO-based detection algorithm, Proposed v11m, that builds on the existing YOLOv11n framework. Our method incorporates a new attention-gate convolution (AGConv) module into the backbone, replacing the original bottleneck to improve feature extraction and cut down unnecessary computations. Additionally, we introduce a multi-dilation sharing convolution (MDSC) module to reduce feature loss during pooling and increase sensitivity to objects of different scales. To further enhance detection accuracy and efficiency, we introduce a lightweight cross-channel feature fusion module (CCFM) within the neck network that dynamically adjusts feature weights to improve multi-scale representation. Experimental results show that the Proposed v11m model achieves a 1.1% increase in [email protected] and a 2.7% improvement in [email protected]:0.95 compared to YOLOv11n, while maintaining real-time performance of 376 FPS with only 2.6 million parameters. These results demonstrate that Proposed v11m offers precise and effective traffic object detection suitable for low-resource devices, even in challenging weather conditions.