4DR360: State Reasoning for Joint 3D Detection and Occupancy Prediction with 4D Radar-Camera Fusion

Summary

4DR360 is a 4D millimeter-wave radar and camera fusion framework for 360-degree full-scene autonomous driving perception, proposed by Xiaokai Bai, Lianqing Zheng, Runwei Guan, Songkai Wang, Siyuan Cao, and Hui-liang Shen (arXiv:2607.09629). The key insight is to model semantic occupancy as a persistent scene state rather than a terminal output. The framework follows a cross-modal state reasoning paradigm with two main components: State-guided BEV Enhancement (SBE), which strengthens intra-frame BEV representations, and Doppler-guided Temporal Fusion (DTF), which preserves state evidence over longer temporal horizons. 4D radar is attractive for autonomous driving because it is robust and cost-effective, but its sparse returns necessitate radar-camera fusion. The authors also extend the ManTruckScenes dataset with a unified cross-dataset detection-occupancy protocol. This paper targets the CV/AutoDrive research area and addresses reliable perception coupling foreground objects with dense semantic layout.