AWARDistill: Adaptive and robust 3D object detection in adverse conditions through knowledge distillation

3D object detection is a crucial component of autonomous vehicle perception, but adverse weather conditions can affect sensor performance, leading to a deterioration in data quality, thereby posing significant challenges to the further development of perception. This paper presents a novel, robust 3D object detection framework to address this issue. Firstly, to tackle the problem of lacking adverse weather datasets, we propose the Multi- modal Adverse-Weather Data Simulation Theory (MIST), which employs optical models to simulate fog and replicates the dynamic properties of rain and snow to recreate real-world circumstances. Secondly, we propose the Adaptive and Robust 3D Object Detection Framework in Adverse Conditions through Knowledge Distillation (AWARDistill), which employs staged knowledge distillation to enable the model to adapt to adverse weather conditions, significantly enhancing detection accuracy and robustness. Additionally, we designed two modules that can be integrated into other detection frameworks to enhance robustness. We evaluated the performance of AWARDistill on multiple datasets. On the KITTI dataset, our model attained an average precision of about 88% and can efficiently adapt to extreme weather. Extensive experiments demonstrate our model's effectiveness and superiority, providing strong support for autonomous driving in challenging weather environments.