PC-YOLO: enhancing object detection in adverse weather through physics-aware and dynamic network structures

Adverse weather conditions, such as rain and fog, significantly impair object detection performance by distorting visual information and reducing image quality. To address this, we introduce PC-YOLO, an advanced object detection model specifically designed to enhance performance in challenging weather scenarios. PC-YOLO integrates two innovative components: the enhanced physics-aware block (EPB) and the conditional width diversification branch block (C2f_WDBB). The EPB leverages physical principles of multi-layer atmospheric scattering to improve feature representation under rainy and foggy conditions while also enhancing channel and spatial features related to the target task. The C2f_WDBB, on the other hand, dynamically adjusts the network width based on weather conditions and target characteristics, enabling the model to adapt to different detection requirements. Experimental evaluations on the real-world task-driven testing set (RTTS), RainyCityscape, and HazyCityscape datasets demonstrate that PC-YOLO outperforms baseline models such as YOLOv11, YOLOv8, faster region-based convolutional neural network (R-CNN), single shot multibox detector (SSD), and RetinaNet. Specifically, PC-YOLO achieves a 2.8% to 20.8% improvement in the mean average precision across datasets, validating the effectiveness of combining physical models with deep learning techniques for robust object detection in complex weather environments. We pave the way for practical applications in autonomous driving and intelligent traffic control.