NA-CWRNet: A Lightweight Lane Detection Method for Complex Weather

The current lane detection methods have shown good performance under high-visibility conditions. However, in complex weather conditions such as fog, rain, and night-time scenarios with low visibility, these methods perform poorly, and in some cases, fail to detect lanes entirely. To address this issue, this paper proposes a lightweight lane detection method that ensures improved detection accuracy and efficiency under such challenging weather conditions. Additionally, the method is designed to be deployable on low-cost, low-computing edge devices, making it suitable for practical applications. Firstly, the image restoration process is introduced to tackle the issue of information loss in low-visibility environments. A dedicated image recovery module, NAFNet, is designed to restore the image before detection. Secondly, a novel detection method is proposed. In response to the limitations of the existing CLRNet method, including high computational cost and a large number of model parameters, we design a lightweight backbone network, MR3-CA, to replace the ResNet backbone in CLRNet. Furthermore, the CAFM attention mechanism is incorporated to integrate both local feature extraction and global feature modeling. This approach effectively captures and models the global and local features of the image, thereby enhancing the detection performance. Experimental results on a self-constructed complex weather dataset show that the proposed method outperforms existing approaches. Specifically, the model's parameters (Params) and computational FLOPs are reduced by 36.2% and 37.6%, respectively. Additionally, the F1 score improves by an average of 27% in foggy conditions, 30% in rainy conditions, and 10% in night-time conditions.