Robust Cross-Modal UAV Detection via Bidirectional Collaborative Energy Field under Weak Calibration

Abstract

The precise detection and localization of 'low-slow-small' (LSS) unmanned aerial vehicles (UAVs) pose significant challenges in terminal defense scenarios. These non-cooperative targets have extremely sparse representations in radar point clouds because of their minimal radar cross-sections and appear as textureless, blurred patches in visual images. Furthermore, traditional multi-modal fusion paradigms rely heavily on rigid geometric projection matrices, which frequently fail under the mechanical vibrations, thermal deformations, and extrinsic perturbations inherent in rapid deployment environments, which often results in severe feature misalignment and contamination. To overcome these critical bottlenecks, we present an end-to-end cross-modal detection framework built on Bidirectional Collaborative Energy Fields (Co-Field). By redefining the 3D object detection task as an energy field reconstruction process within a latent multimodal feature space, we entirely remove the need for absolutely precise extrinsic parameters. Our framework uses a novel bidirectional mutual-guided interaction mechanism. Specifically, a vision-guided radar branch employs deformable cross-attention to adaptively aggregate highly discrete radar geometric features around visual semantic centers, achieving a feature-level soft alignment that implicitly compensates for calibration drift. Concurrently, a radar-corrected vision branch utilizes radar-derived geometric occupancy probabilities as depth anchors to gate and calibrate visual features, effectively suppressing the inherent divergence of monocular depth estimation. Ultimately, the framework regresses the target's precise 3D states and fine-grained categories directly from the extrema of the reconstructed collaborative energy field, circumventing complex heuristic post-processing. Experiments on the curated MMAUD dataset show that the proposed Co-Field method achieves better robustness. It remains highly effective even under severe calibration errors up to 10 degrees and achieves a 74.2% recall rate at extreme sparse distances ranging from 70 to 100 meters, and consistently outperforms state-of-the-art rigid and soft fusion baselines.