Multiscale Neighborhood Cluster Scene Flow Prior for LiDAR Point Clouds

Scene flow estimation, which aims to predict point-wise displacement in 3-D space from sequential data, is a challenging task with wide application in fields such as robotics and autonomous driving. Currently, the accuracy of scene flow estimation from sparse point clouds using prior-based models is suboptimal. Therefore, we revisit the point-by-point scene flow prior and propose a multiscale neighborhood cluster scene flow prior (MNCSFP) to enhance the accuracy of scene flow estimation in sparse point clouds. We optimize the prior model utilizing the multiscale neighborhood cluster feature of the point cloud, in which the point neighborhood is constructed only once. According to the neighborhood index (NI), we design a multiscale neighborhood cluster feature construction (MNCFC) module. The MNCFC module rapidly constructs multiscale neighborhood cluster features using a Gaussian-based neighborhood feature normalization (GNFN) strategy to improve the representation of neighborhood cluster characteristics. Moreover, we propose a neighborhood cluster weighted aggregation (NCWA) module to encode neighborhood cluster features. In NCWA, we design the logsoft function to calculate the neighborhood cluster weights and complete the extraction and aggregation of neighborhood cluster features. Furthermore, we design the multiscale feature fusion (MFF) module that combines the symmetry operation and the logsoft function to complete the fusion of multiscale features to enhance the sparse point feature stability. We evaluate our method on samples from the KITTI, Argoverse, nuScenes, and Waymo Open datasets and demonstrate that it outperforms existing methods and achieves advanced performance.