Source-linked AI summary
Intensity Scan Context: Coding Intensity and Geometry Relations for Loop Closure Detection
Han Wang, Chen Wang, Lihua Xie
TL;DR
The paper tackles LiDAR loop closure detection that largely relies on geometric descriptors while ignoring intensity information. It proposes intensity scan context with a two-stage hierarchical retrieval strategy, and reports higher recall precision and recall rate than geometric-only LiDAR methods, with 1.2 ms per query.
Problem
Existing 3D loop closure methods often use geometric-only descriptors and leave LiDAR intensity information unused, despite intensity potentially distinguishing places.
Method
The method introduces intensity scan context, a global descriptor integrating geometry and intensity, with binary-operation-based geometry retrieval followed by intensity structure matching.
Results
The method achieves higher recall precision and recall rate than LiDAR-based geometric-only approaches across three public datasets, at 1.2 ms per query.
Takeaways & Limitations
LiDAR intensity can be effective for place recognition, while the proposed hierarchical ISC retrieval is reported as efficient in practice.
Abstract
from arXiv · showhide
Loop closure detection is an essential and challenging problem in simultaneous localization and mapping (SLAM). It is often tackled with light detection and ranging (LiDAR) sensor due to its view-point and illumination invariant properties. Existing works on 3D loop closure detection often leverage the matching of local or global geometrical-only descriptors, but without considering the intensity reading. In this paper we explore the intensity property from LiDAR scan and show that it can be effective for place recognition. Concretely, we propose a novel global descriptor, intensity scan context (ISC), that explores both geometry and intensity characteristics. To improve the efficiency for loop closure detection, an efficient two-stage hierarchical re-identification process is proposed, including a binary-operation based fast geometric relation retrieval and an intensity structure re-identification. Thorough experiments including both local experiment and public datasets test have been conducted to evaluate the performance of the proposed method. Our method achieves higher recall rate and recall precision than existing geometric-only methods.
I. INTRODUCTION
The paper addresses LiDAR loop closure detection by adding intensity information to geometric descriptors. It proposes ISC and a two-stage retrieval strategy, evaluated on local and public datasets.
- Motivation: LiDAR supports place recognition because it is less affected by illumination, weather, and viewing-angle changes than vision.Loop closure detection can reduce trajectory drift and prevent duplicate landmark registration in SLAM.
- Motivation: Existing 3D descriptors use global or local geometry while leaving LiDAR intensity information unused.Intensity is affected by surface characteristics, acquisition geometry, and instrument effects, making it less straightforward to exploit.
- Contributions: Intensity scan context (ISC) is a global descriptor that integrates geometry and intensity characteristics for loop closure detection.The authors argue that reflectance structure can be distinctive across places.
- Contributions: The proposed two-stage hierarchical retrieval combines binary-operation-based fast geometry retrieval with intensity structure matching.The strategy is designed to improve loop closure detection efficiency.
- Evaluation: 1.2 ms per query is reported for the proposed identification process.The paper evaluates the descriptor through local experiments and public-dataset tests.
II. RELATED WORK
Related work spans vision-based and LiDAR-based place recognition, with LiDAR offering robustness but geometric descriptors facing efficiency or competitiveness limits. Prior intensity-aware work motivates a more efficient global descriptor.
- Vision-based methods: Vision-based methods commonly use bag-of-words models, but image streams are vulnerable to illumination and viewpoint changes.Examples include FAB-MAP and DBoW2 in visual SLAM systems.
- LiDAR-based methods: LiDAR was introduced for loop closure detection because of its robustness to illumination and viewpoint changes.LiDAR-based methods seek efficient local descriptors or concise global place signatures.
- Descriptor efficiency: Local-descriptor approaches can require key-point extraction and extensive geometry calculations, whereas global descriptors are more efficient for place recognition.GLAROT encodes relative geometric positions, while Scan Context projects laser scans into a global descriptor.
- Intensity-aware methods: Prior intensity-aware work showed improved performance over geometric-only descriptors but remained computationally expensive.ISHOT combines geometry and intensity within a probabilistic voting strategy.
III. METHODOLOGY
The methodology calibrates and preprocesses LiDAR intensity before constructing an intensity scan context from geometry-defined subspaces. The resulting representation preserves both environmental geometry and intensity structure.
- Intensity calibration and pre-processing: LiDAR intensity estimates surface reflectance from returned energy, and readings vary across object materials.Examples include high returns from retro-reflective metal and lower returns from concrete.
- Intensity scan context: Intensity scan context projects intensity onto subspaces formed by decomposing the original point cloud according to geometry characteristics.The resulting global signature represents the environment as a 2D matrix containing geometry and intensity distribution.
- Intensity calibration and pre-processing: Intensity readings require calibration because surface properties, acquisition geometry, and instrument effects all affect the measured signal.The distance influence is modeled using a mapping function ϕ.
- Intensity calibration and pre-processing: Preprocessing removes unreliable distant points using a distance threshold and reduces the influence of ground points.The paper notes that LiDAR noise increases with distance.
B. Intensity Scan Context
Intensity scan context represents a LiDAR scan as a global 2D signature that combines geometry and intensity information. It partitions the scan into polar subspaces and codes each subspace’s intensity to reveal environmental structure.
- Intensity scan context integrates geometry and intensity characteristics into an efficient global signature for 3D LiDAR scans.
- Each LiDAR point contains Cartesian geometry and intensity, with the scan converted to polar coordinates in the x-y plane.
- The point cloud is divided into Ns sectors and Nr rings using equally spaced azimuthal and radial partitions.
- Each polar subspace applies a coding function κ to reduce its intensity dimension because intensity is assumed to vary little within small regions.
- Empty subspaces receive ηij = 0, and the resulting global signature Ω is a 2D matrix encoding the environment’s geometry and intensity distribution.
C. Place Re-identification
Place re-identification matches a current LiDAR scan against previously visited scans in a growing historical database. The proposed retrieval strategy uses fast binary operations to reduce the increasing computational cost.
- Place recognition matches the current place Pn with previously visited places in the historical database D = {P1, P2, · · · , Pn−1}.
- As more places are visited, the database grows and increases the computational cost of place re-identification.
- A two-stage hierarchical intensity scan context retrieval strategy uses fast binary operations to accelerate place re-identification.
1) Fast Geometry Re-identification:
Fast geometry re-identification converts intensity scan contexts to binary matrices and compares them with XOR operations. Column shifts account for viewpoint changes, while a threshold filters unmatched pairs.
- Fast Geometry Re-identification: Binary-operation geometry re-identification uses binary transforms of intensity scan contexts for fast indexing.
- Fast Geometry Re-identification: Geometry similarity is computed from the proportion of unequal elements obtained through element-wise XOR between query and candidate binary matrices.
- Fast Geometry Re-identification: Because columns represent azimuthal directions, laser-scan rotations are handled as column shifts of the intensity scan context.
- Fast Geometry Re-identification: The method identifies the best matching column shift, which can be used to correct the viewing-angle change.
- Fast Geometry Re-identification: 0.5 ms is the reported binary-matching cost on a desktop computer, and pairs below an empirically determined threshold ϵg can be filtered out.
2) Intensity Structure Matching:
Intensity structure matching compares query and candidate intensity scan contexts column by column using cosine distance. It reuses the geometry-derived column shift and applies an empirical threshold to reject unmatched pairs.
- Intensity Structure Matching: The second retrieval stage identifies intensity similarity between query and candidate intensity scan contexts through column-wise comparison.
- Intensity Structure Matching: Column-wise intensity similarity is measured using cosine distance.
- Intensity Structure Matching: The geometry stage’s identified column shift is reused to correct the viewing-angle change before comparing intensity structures.
- Intensity Structure Matching: An empirically determined threshold ϵi filters unmatched pairs after intensity matching.
D. Consistency Verification
Because a global descriptor simplifies the original point cloud, it can ignore features and produce false positives, so consistency must be checked before loop closure.
- Global descriptors may ignore features because they simplify the original point cloud.
- Consistency checking is necessary before accepting a loop closure to address possible false positives.
1) Temporal consistency check:
The method verifies candidate loop closures using temporal consistency across neighbouring LiDAR scans, with special handling for reverse visits and a final acceptance threshold.
- Temporal consistency check:: A single loop closure implies high similarity among neighbouring LiDAR scans because sensor feedback is continuous in time.
- Temporal consistency check:: Temporal consistency uses N frames to verify candidate loop closures.
- Temporal consistency check:: For reverse visits, the temporal comparison changes from I_m−i to I_m+i after geometry detects a viewing-angle change.
- Temporal consistency check:: A candidate loop is accepted when its final temporal consistency score meets threshold ξ.
A. Experiment Setup
Experiments evaluate the proposed method in an indoor warehouse and on KITTI sequences, comparing loop-closure detection against vision- and LiDAR-based methods using recall and precision.
- A. Experiment Setup: The indoor experiment uses an autonomous guided vehicle with Velodyne VLP-16 and Intel Realsense r200 in a warehouse environment.
- A. Experiment Setup: The warehouse task includes leaving a docking station, fetching materials, and returning in reverse direction.
- B. Experiment on Autonomous Robot: The proposed method identifies the revisited warehouse place, while DBoW2 fails under the reverse-visit viewpoint change and produces an incorrect trajectory.
- C. Evaluation on Public Dataset: KITTI evaluation tests sequences 00, 02, and 05, with sequence 02 containing both forward and reverse visits.
- C. Evaluation on Public Dataset: Recall rate and precision are collected for each test, with GPS used to determine the total number of loop closures.
- C. Evaluation on Public Dataset: Compared with LiDAR-based methods, the approach achieves higher recall precision and recall rate across all three datasets, at 1.2 ms/query.
V. CONCLUSION
The paper integrates LiDAR geometry and intensity in ISC and uses hierarchical retrieval to reduce computational cost. Experiments report competitive recall precision and recall rate against state-of-the-art methods.
- The proposed loop-closure approach integrates geometry and intensity information in a global 3D descriptor named intensity scan context.
- The two-stage hierarchical retrieval combines binary-operation-based geometry indexing with intensity structure re-identification.
- The method costs 1.2 ms per query on a normal computer in practice.
- Experiments include an autonomous warehouse robot and public datasets, with competitive recall precision and recall rate compared to state-of-the-art methods.