【摘要】 动态场景的高精度重建在自动驾驶、增强现实等领域具有重要意义。本文提出一种基于事件相机的动态神经辐射场重建与运动补偿方法 STAA-NeRF。该方法设计可微分事件熵引导的自适应时空注意力模块（DEE-SA），在多帧事件流中自适应聚合动态特征，提升对高速运动区域的建模能力；同时引入事件-射线微分耦合模型（ERDCM），通过事件梯度驱动实现采样偏移与密度调节，从而在理论上保证运动补偿与几何建模的一致性。实验结果表明，在真实及合成的高速场景中，STAA-NeRF相较于Deblur-NeRF平均提升约2.6 dB PSNR、0.1SSIM，及降低0.04 LPIPS，在低光照与高速运动环境下依然保持鲁棒性能。研究成果为动态视觉系统的实时感知与高质量重建提供了一条新技术路径。更多还原

【Abstract】 High-precision reconstruction of dynamic scenes is of great significance in fields such as autonomous driving and augmented reality. This paper proposes a dynamic Neural Radiance Field reconstruction and motion compensation method based on event cameras, called STAA-NeRF. This method designs a differentiable event entropy-guided adaptive spatio-temporal attention module(DEE-SA), which can adaptively aggregate dynamic features in multiple frames of event streams, improving the modeling ability for high-speed moving areas. At the same time, it introduces an event-ray differential coupling model(ERDCM), which realizes sampling offset and density adjustment through event gradients, thereby theoretically ensuring the consistency of motion compensation and geometric modeling. Experimental results show that in real and synthetic high-speed scenes, STAA-NeRF outperforms Deblur-NeRF with a 2.6 dB increase in PSNR, a 0.1 increase in SSIM, and a 0.04 reduction in LPIPS, and still maintains robust performance in low-light and high-speed motion environments. The research results provide a new technical path for real-time perception and high-quality reconstruction of dynamic vision systems.更多还原