LiDAR Inertial Odometry And Mapping Using Learned Registration-Relevant Features

TL;DR

This paper proposes a learning-based approach that select points relevant to LiDAR SLAM pointcloud registration that performs well on multiple public benchmarks and demonstrates the effectiveness of the learning-based feature extraction module through comparison with several handcrafted feature extractors.

Abstract

SLAM is an important capability for many autonomous systems, and modern LiDAR-based methods offer promising performance. However, for long duration missions, existing works that either operate directly the full pointclouds or on extracted features face key tradeoffs in accuracy and computational efficiency (e.g., memory consumption). To address these issues, this paper presents DFLIOM with several key innovations. Unlike previous methods that rely on handcrafted heuristics and hand-tuned parameters for feature extraction, we propose a learning-based approach that select points relevant to LiDAR SLAM pointcloud registration. Furthermore, we extend our prior work DLIOM with the learned feature extractor and observe our method enables similar or even better localization performance using only about 20\% of the points in the dense point clouds. We demonstrate that DFLIOM performs well on multiple public benchmarks, achieving a 2.4\% decrease in localization error and 57.5\% decrease in memory usage compared to state-of-the-art methods (DLIOM). Although extracting features with the proposed network requires extra time, it is offset by the faster processing time downstream, thus maintaining real-time performance using 20Hz LiDAR on our hardware setup. The effectiveness of our learning-based feature extraction module is further demonstrated through comparison with several handcrafted feature extractors.

Figures (6)

• Figure 1: Accurate and detailed map produced by DFLIOM on Newer College Short (1.4 km) colored by intensity. Our method provides accurate trajectory (yellow) and map while significantly decreasing memory usage. Zoom-in showcase the details DFLIOM captures.
• Figure 2: Architecture of the proposed feature extraction network. Kernel Point Convolution (KPConv) is used as the backbone for extracting a higher dimensional representation from point clusters. Inspired by PointNet we use shared MLPs to map point coordinates to higher dimension as positional embedding. Separate shared MLPs are used to predict the saliency and uniqueness scores, based on which more important points are chosen. Shared weights make our network sufficiently light-weight to inference in real-time.
• Figure 3: Example point clouds after feature extraction. blue: selected feature points. (a): When only selecting best Saliency Features, parallel walls are selected, which can be featureless and similar to neighboring scans. (b): When only selecting best Unique Features, detailed features near the robot are selected and thus only useful for local scale scan-to-scan matching. (c): With both types of features, the extracted point clouds are rich in feature and provides good coverage.
• Figure 4: Example point clouds (colored by intensity) used by our method. (a): Point cloud after feature extraction (white) (b): Feature extraction is not performed when the robot is in a narrow corridor.
• Figure 5: Memory usage (RSS in MB) vs. time along trajectory on Newer College Dataset ramezani2020newer Short with DFLIOM (green) and DLIOM (orange). DFLIOM uses significantly less memory (e.g., from 14.6GB to 5.5GB) while maintaining localization accuracy. Our Python feature extractor (blue) uses constant RAM ($\sim$1GB).