Lightweight multiscale object detection for complex weather conditions

Abstract

To address the challenges of object detection in complex weather conditions, including occlusion, blurred boundaries, and small object perception, this paper proposes a lightweight detection model based on YOLOv11n. First, PP-LCNet is adopted as the backbone to reduce the parameter size and computational complexity. Second, a multidimensional collaborative attention module named W-MCA is introduced to enhance channel and spatial feature representation during feature fusion. Third, an adaptive regression loss function called LA-MPDIoU is designed by incorporating center-distance and diagonal-variance terms with learnable weighting to improve localization accuracy. Experimental results on the BDD100K dataset demonstrate that the proposed model achieves a precision of 62.3 percent and a mean average precision of 50.5 percent, showing absolute improvements of several percent over the baseline while using only 2.4 million parameters. These results indicate that the proposed method effectively improves detection accuracy while maintaining a lightweight model suitable for deployment in complex environments.