Motion-aware dynamic feature association for asynchronous multi-modal 3D object detection

Multi-modal 3D object detection typically assumes strict temporal synchronization between cameras and LiDAR. However, sensor streams are often asynchronous in practice, causing cross-modal observations to correspond to different physical states of dynamic objects. Existing fusion methods either ignore this misalignment or rely on implicit warping, both of which degrade significantly under realistic temporal offsets. To address this challenge, we propose MADFA, a motion-aware framework for asynchronous multi-modal 3D detection. Specifically, we introduce Confidence-Adaptive Motion Bias (CAMB), a cross-attention mechanism whose spatial support is adaptively governed by a motion reliability score to handle unreliable motion cues. Furthermore, we derive an identifiability bound for cross-modal recoverability under asynchronous sensing, which directly guides the design of CAMB. In tandem, a self-supervised Motion Cue Reliability (MCR) score is formulated to jointly model LiDAR sparsity and image tracking consistency. Comprehensive experiments on nuScenes and Waymo under multiple asynchrony protocols demonstrate that MADFA consistently outperforms competitive baselines. Moreover, our framework exhibits graceful degradation as offsets increase, fully corroborating the theoretical recoverability condition.