Random Walk on Point Clouds for Feature Detection
This study addresses the challenge of keypoint detection in point clouds by proposing a two-stage context-aware approach. It first introduces a Disk Sampling Neighborhood (DSN) descriptor designed to preserve local neighborhood structure, then constructs a graph that integrates spatial distribution, topological relationships, and geometric features. On this unified graph representation, a random walk strategy (RWoDSN) is employed to identify salient keypoints. Notably, this work is the first to jointly incorporate these three complementary sources of information within a graph-based random walk framework, thereby overcoming limitations inherent in conventional geometry-only invariant methods. Experimental results demonstrate significant performance gains across eight evaluation metrics, achieving a recall of 0.769—representing up to a 22% improvement—and a precision of 0.784, outperforming existing state-of-the-art techniques.