PF-HDiffBEV: Bird’s eye view feature enhancement mechanism based on potential field-guided hierarchical conditional diffusion model for autonomous driving

Bird’s eye view (BEV) representation provides a new paradigm for environmental perception in autonomous vehicles, and it has been widely adopted for 3D object detection. However, when structural factors, such as object occlusion, distant objects, or complex backgrounds, lead to information loss, BEV features still have difficulty effectively extracting and representing scene information, which degrades detection performance. To address the demanding engineering problem of achieving high-precision 3D object detection under conditions with missing scene information, this paper proposes a BEV feature enhancement mechanism based on a potential field (PF)-guided hierarchical conditional diffusion model (PF-HDiffBEV). This mechanism is integrated as a universal paradigm into different BEV frameworks to improve detection performance. First, physical priors of scene structure are explicitly encoded via PF modeling, and the resulting PF distribution is used as the diffusion model’s guidance target. Subsequently, the original BEV features with missing scene information are used as conditions, and hierarchical conditional embedding is applied to denoise and reconstruct the predicted PF features from Gaussian white noise. Finally, residual networks and channel concatenation fuse the predicted PF features with the original BEV features to produce enhanced BEV features that improve 3D object detection performance. These enhanced features are then provided to the object detection task head. Experimental results show that PF-HDiffBEV achieves significant and stable performance gains across three representative BEV 3D object detection frameworks: camera, multi-sensor fusion, and temporal fusion. These findings confirm PF-HDiffBEV’s effectiveness and advanced capability in explicitly compensating for missing scene information and strengthening BEV feature representation while demonstrating strong cross-architecture generalization. PF-HDiffBEV also shows robust detection performance and real-time computational capability in real-world field tests, satisfying engineering requirements for real-vehicle 3D object detection.