LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis

LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis

2024-04-03 | Zehan Zheng, Fan Lu, Weiyi Xue, Guang Chen†, Changjun Jiang
LiDAR4D is a novel framework for dynamic space-time LiDAR view synthesis, addressing the challenges of dynamic reconstruction, large-scale scene characterization, and realistic synthesis. The method introduces a 4D hybrid representation combining multi-planar and grid features to effectively reconstruct LiDAR point clouds in a coarse-to-fine manner. Geometric constraints derived from point clouds are incorporated to improve temporal consistency, and global optimization of ray-drop probability is used to preserve cross-region patterns. Extensive experiments on KITTI-360 and NuScenes datasets demonstrate the superior performance of LiDAR4D in geometry-aware and time-consistent dynamic reconstruction. The method outperforms previous state-of-the-art approaches in terms of Chamfer Distance error, range depth, and intensity metrics. LiDAR4D's effectiveness is further validated through ablation studies and additional experiments, showcasing its adaptability and potential for various applications in autonomous driving and robotics.LiDAR4D is a novel framework for dynamic space-time LiDAR view synthesis, addressing the challenges of dynamic reconstruction, large-scale scene characterization, and realistic synthesis. The method introduces a 4D hybrid representation combining multi-planar and grid features to effectively reconstruct LiDAR point clouds in a coarse-to-fine manner. Geometric constraints derived from point clouds are incorporated to improve temporal consistency, and global optimization of ray-drop probability is used to preserve cross-region patterns. Extensive experiments on KITTI-360 and NuScenes datasets demonstrate the superior performance of LiDAR4D in geometry-aware and time-consistent dynamic reconstruction. The method outperforms previous state-of-the-art approaches in terms of Chamfer Distance error, range depth, and intensity metrics. LiDAR4D's effectiveness is further validated through ablation studies and additional experiments, showcasing its adaptability and potential for various applications in autonomous driving and robotics.
Reach us at info@study.space
[slides and audio] LiDAR4D%3A Dynamic Neural Fields for Novel Space-Time View LiDAR Synthesis