Leveraging contrastive predictive coding to enhance radar-based object detection under adverse weather conditions

Robust object detection is essential for safe autonomous driving and is performed using cameras, LiDAR, and radar sensors. In adverse weather conditions, radar offers a clear advantage as its longer operating wavelength is less affected by rain, fog, and snow than RGB cameras and LiDAR. However, radar-only object detection remains challenging due to the noisy and low-resolution nature of radar data. This paper presents a radar-only object detection framework that incorporates a self-supervised temporal predictive learning module inspired by contrastive predictive coding (CPC) to extract temporal representations from unlabeled radar sequences. The proposed method adapts the predictive learning principle of CPC to 3D radar tensors in a lightweight form suitable for high-dimensional radar data. The temporal encoder is pretrained on radar sequences to produce feature embeddings that capture inter-frame dependencies. These learned temporal features are then integrated into a modified K-Radar baseline network. Experiments on the K-Radar dataset show that the proposed framework improves radar-only detection performance under adverse weather conditions. In heavy snow, for example, the proposed method achieves mean average precision values of 80.73% and 69.44% for the sedan and bus/truck classes, respectively, compared with 71.46% and 59.63% for the baseline radar model. Comparisons with a temporal representation learning approach based on TS2Vec further show that the proposed method achieves superior overall detection accuracy, while TS2Vec remains a competitive lightweight alternative.