All-day perception for intelligent vehicles: switching perception algorithms based on WBCNet

A weather- and brightness-based classification network (WBCNet) is proposed for driving scene classification to address the decreased accuracy in perception caused by weather and environment changes. To facilitate its applicability in vehicles and minimize computational demands on vehicle chips, WBCNet has been designed with special modules, including attention mechanisms and dilated convolutions. Dilated convolutions combined with residual connections empower WBCNet to concurrently handle information at various scales. This aids in simplifying the training and optimization of deep networks, consequently enhancing the model's performance and mitigating the risk of overfitting. The outstanding feature association capability originating from the fusion of channel attention and spatial attention enables WBCNet to focus more on the sky, lanes, and other traffic information features within the image. This design enables WBCNet to use only images as input, making it highly suitable for engineering applications. The output of WBCNet provides the basis for the downstream perception model selection algorithm, allowing it to choose the appropriate perception model for different scenes accurately. A dataset with complex scenes based on Carla is constructed for comparison to verify WBCNet's performance. Finally, a real-world driving dataset is used to validate the effectiveness and real-time performance of WBCNet.