Vehicle and Pedestrian Detection in Extreme Weather Based on Improved Rt-Detr

Adverse weather conditions often lead to the loss or blurring of visual information, which affects the accuracy of object detection. To improve the accuracy of vehicle and pedestrian detection under such conditions, this paper proposes an improved method based on RT-DETR. Firstly, the DySample Triple Feature Encoder module (DS-TFE) is designed to replace the Concat module in RT-DETR, this modification enhances the efficiency of multi-scale feature processing Secondly, the Adaptive Spatial Feature Fusion Module (ASFF) is introduced with the objective of reducing semantic conflicts between different feature maps in RT-DETR. Prior to the commencement of the experiment, the background noise present within the image is reduced through the utilisation of an optimized Prior of the Dark and Bright Channel (PDA). The experimental outcomes demonstrate that the proposed improved model performs well under a multitude of extreme weather conditions, particularly under conditions of heavy rain and fog, achieving enhanced detection accuracy whilst maintaining a low computational (79.5 GFLOPs). The results demonstrate the effectiveness of this method under extreme weather conditions.