객체 분할을 위한 가변적 반경 기반의 3D LiDAR 포인트 클라우드 그룹핑

Since a 3D point cloud acquired by LiDAR sensors is a discrete set of data points in space, its dimensionality and sparsity are commonly high which limit its applicability. An intuitive yet convincing resolution to overcome these challenges is to first group the data points and then segment objects out for subsequent processes. Among existing methods in the literature, density-based clustering has been among the popular choices due to its promising performance. However, the applicability of the density-based approach is limited due to its hyperparameter, a distance between data points. In this work, we propose an effective and hyperparameter-free method for grouping LiDAR points. Noise points such as ground clutters are first filtered out by the proposed cone-shaped filtering. The remaining points are then grouped as object candidates by the proposed distance variation analysis. From our experimental study, the proposed method is proven to better segment objects out than those of the compared methods.