Robust 3D object detection in challenging weather scenarios remains a significant challenge due to sensor and algorithm degradation caused by various environmental noises. In this paper, we propose a novel camera-radar-based 3D object detection framework that leverages a cross-modality knowledge distillation method to improve detection accuracy in adverse conditions, such as rain and snow. Specifically, we introduce a teacher-student training paradigm, where the teacher model is trained under clear weather and guides the student model trained under weather-degraded environments. We design three novel distillation losses focusing on spatial alignment, semantic consistency, and prediction refinement between different modalities to facilitate effective knowledge transfer. Moreover, a weather simulation module is introduced to generate adverse-weather-like input, enabling the student model to learn robust features under challenging conditions better. A gated fusion module is also integrated to adaptively fuse camera and radar features, enhancing robustness to modality-specific degradation. Experimental results on the nuScenes dataset reveal our model outperforms multiple state-of-the-art methods, achieving superior results across common detection metrics (mAP, NDS) and per-class AP, particularly under challenging weather, showing improvements of 3.5–3.9 % mAP and 4.3–4.8 % NDS in rainy and snowy scenes.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号