SwinCGH-Net: Enhancing Robustness of Object Detection in Autonomous Driving with Weather Noise via Attention

Abstract

Object detection in autonomous driving requires high accuracy and speed in different weather. At present, many CNN-based networks have achieved high accuracy on academic datasets, but their performance disastrously degrade when images contain various kinds of noises, which is fatal for autonomous driving. In this paper, we propose a detection network based on shifted windows Transformer (Swin Transformer) called SwinCGH-Net, with a kind of new detector head based on lightweight convolution attention module, which makes full use of the attention mechanism in both feature extraction and detection stages. Specifically, we use Swin Transformer as backbone to extract feature in order to obtain effective information from a small amount of pixels as well as integrate global information. Then we further improve the robustness of the network through the detector head contained lightweight attention block S-CBAM. Furthermore, we use Generalized Focal Loss to calculate loss, which effectively enhances the representation ability of the model. Experiments on Cityscapes and Cityscapes-C datasets demonstrate the superiority and effectiveness of our method in different weather condition. With the increasing level of weather noise, our method shows strong robustness compared with previous method, especially in small object detection.