ADA-DPM: A neural descriptor-based SLAM strategy for coarse-to-fine point cloud registration

Abstract

Simultaneous Localization and Mapping (SLAM) is a fundamental technique for pose estimation and map construction in unknown environments. However, existing approaches often struggle to balance localization accuracy and system robustness, particularly in dynamic, noisy, and unstructured scenes. To address these challenges, we propose ADA-DPM, an adaptive point-cloud noise filtering SLAM strategy based on coarse-to-fine registration. First, to mitigate the influence of noise and distorted points, we introduce an Importance Scoring Head, which dynamically selects registration-informative feature points according to coarse registration and predicted importance scores. This head emphasizes reliable correspondences while suppressing low-importance points that may correspond to noise or outliers. Second, we design a Dynamic Segmentation Head to remove dynamic points and reduce the interference caused by moving objects and ego-motion errors. Furthermore, we introduce the Cross-Layer Intra-Graph Convolution Module (CLI-GCN), which samples shallow-layer spatial neighborhoods for deep-layer feature points and constructs local neighborhood graphs to integrate multi-scale geometric information, thereby improving feature representation in overlapping regions. Experiments on multiple public datasets demonstrate that ADA-DPM achieves competitive performance in challenging real-world scenarios, confirming its effectiveness in both localization accuracy and system robustness.