ADA-DPM: A Neural Descriptors-based Adaptive Noise Filtering Strategy for SLAM

TL;DR

ADA-DPM tackles dynamic object interference, noise, and loop-closure reliability in LiDAR SLAM by introducing three core modules: Dynamic Segmentation Head to remove dynamic points, Global Importance Scoring Head to weight high-contribution feature pairs, and GLI-GCN to fuse multi-scale geometric information. The encoder–decoder framework embeds PointNeXt and CLI-GCN, with a quaternion-based rotation prediction and translation recovered from centroids, all optimized via a suite of loss terms including InfoNCE, Mahalanobis-based offset, and an importance-driven reconstruction loss. Experimental results across SemanticKITTI, KITTI-360, and MulRan demonstrate improved localization accuracy, robustness to noise, and reduced map memory, with ablations confirming the value of each module. The work advances robust, adaptive LiDAR-SLAM capable of handling dynamic scenes and complex loop-closure regions, albeit with higher runtime than purely geometric methods and a need for richer annotations in some regimes.

Abstract

Lidar SLAM plays a significant role in mobile robot navigation and high-definition map construction. However, existing methods often face a trade-off between localization accuracy and system robustness in scenarios with a high proportion of dynamic objects, point cloud distortion, and unstructured environments. To address this issue, we propose a neural descriptors-based adaptive noise filtering strategy for SLAM, named ADA-DPM, which improves the performance of localization and mapping tasks through three key technical innovations. Firstly, to tackle dynamic object interference, we design the Dynamic Segmentation Head to predict and filter out dynamic feature points, eliminating the ego-motion interference caused by dynamic objects. Secondly, to mitigate the impact of noise and unstructured feature points, we propose the Global Importance Scoring Head that adaptively selects high-contribution feature points while suppressing the influence of noise and unstructured feature points. Moreover, we introduce the Cross-Layer Graph Convolution Module (GLI-GCN) to construct multi-scale neighborhood graphs, fusing local structural information across different scales and improving the discriminative power of overlapping features. Finally, experimental validations on multiple public datasets confirm the effectiveness of ADA-DPM.

Figures (10)

• Figure 1: The SLAM pipeline of ADA-DPM. The green part represents the odometry part; the blue part represents loop closure detection; the white part represents pose graph optimization; the blue-boxed portion within the white part represents the final global map and trajectory obtained.
• Figure 2: The network overview of ADA-DPM. The blue part represents the Encoder, and the white part represents the Decoder. The networks for both the source point cloud and the target point cloud share the same weights.
• Figure 3: Overview of Dynamic Segmentation Head and Importance Scoring Head. (a) is the overview of Dynamic Segmentation Head; (b) is the overview of Importance Scoring Head.
• Figure 4: Overview of CLI-GCN. (a) is the overview of Cross-Layer Feature Point Sample; (b) is the the overview of Intra-Graph Convolution.
• Figure 5: Visualization results of deep feature points and sampled local neighborhood graph feature points. The yellow point denotes the shallow feature point; the blue point denotes the deep feature point.