A solid-state lidar-inertial 3D scene reconstruction method based on wearable laser helmet

Changjun Gu, Jiaxu Leng, ZongYi Xu, XinBo GaoScientia 2024 · DOI

TL;DR

Researchers have developed a robust 3D scene reconstruction method specifically for wearable laser helmets. By introducing a motion compensation technique based on intermediate reference points and a feature-consistency weighting model, the system overcomes the extreme jitter and rapid turns typical of human head movement, outperforming existing state-of-the-art methods in complex environments like forests and heritage sites.

What Problem Does This Paper Address?

3D reconstruction is a cornerstone of autonomous driving and resource exploration. While vehicle-mounted or drone-based LiDAR (Light Detection and Ranging) systems are common, they are limited by where they can travel. Wearable sensors—like a helmet-mounted system—offer the ultimate flexibility for mapping GNSS-denied or tight spaces like caves, dense forests, or indoor tunnels.

However, humans are "noisy" platforms. Unlike a smooth-rolling car, a person’s head undergoes constant, high-frequency jitter and sudden, sharp rotations. Standard LiDAR-Inertial Odometry (LIO) algorithms often struggle with:

  1. Motion Distortion: Because LiDAR scans points sequentially, rapid head movement "smears" the data within a single frame.
  2. Feature Matching Failure: In non-structured environments (like a forest), traditional algorithms struggle to find the clean geometric planes they need to calculate position, leading to "drift" or total system failure.

Methods

The authors propose a system combining a solid-state LiDAR and an Inertial Measurement Unit (IMU) using an Iterative Kalman Filter (IKF) framework. Their novel contributions focus on two areas:

  • Intermediate Reference Point Compensation: Most standard methods project all laser points in a frame to the position at the end of the scan to correct for motion. This paper argues that projecting to the middle of the scan interval minimizes the cumulative error from the IMU, leading to significantly higher precision during rapid turns.
  • Local Neighborhood Feature Consistency: Not all geometric features are equally reliable. The authors developed a weighting model that evaluates the "consistency" of local planes. If a neighborhood of points has highly aligned surface normals (like a flat floor), it receives a high weight. If the normals are chaotic (like tree leaves), the weight is lowered. This prevents "noisy" features from pulling the trajectory estimate off-course.

Datasets

The study evaluated the algorithm using:

  1. WHU-Helmet Dataset: A public benchmark containing five distinct scenarios: Roadside (urban), Mountain (forest), Tunnel (underground), Infrastructure (structured), and Heritage (complex architecture).
  2. Custom Campus Dataset: The researchers built their own wearable helmet (using a Livox-Avia LiDAR and XSENS-300 IMU) to capture data across a university campus, including teaching buildings, football fields, and wooded areas.

Key Insights and Ideas

  • The "Middle" Ground Matters: The shift from end-of-frame to middle-of-frame projection is a simple but powerful geometric insight. By halving the maximum time-offset for any given point, they effectively halved the window for IMU drift to manifest as spatial distortion.
  • Robustness through Selective Listening: Traditional LIO often treats all detected "planes" with equal importance. By implementing a consistency check, the algorithm essentially "listens" more to reliable structures (walls, floors) and "ignores" the chaotic geometry of vegetation, which is the primary cause of failure in forest mapping.
  • Solid-State Advantage: By using solid-state LiDAR rather than traditional mechanical spinning units, the system is lighter, cheaper, and more durable—essential for a device meant to be worn on a human head for extended periods.

Key Results

The proposed method was compared against industry standards like Fast-LIO2 and Livox-Mapping.

  • Accuracy: In the "Roadside" scenario, the proposed method achieved a Root Mean Square Error (RMSE) of 4.794m, compared to 5.724m for Fast-LIO2.
  • Survivability: The most striking result occurred in the "Heritage" and "Tunnel" datasets. While all competing methods (Fast-LIO2 and Livox-Mapping) failed completely to reconstruct the Heritage site due to its complex, close-range geometry, the authors' method succeeded with an RMSE of 7.540m.
  • Qualitative Detail: Visual results show much sharper reconstructions of stairs and narrow corridors, where motion blur usually ruins the map.

Limitations and Future Work

While the system handles "noisy" human motion well, it still faces challenges in extremely featureless environments. For example, in perfectly smooth, long tunnels, the lack of geometric "anchors" can still lead to longitudinal drift.

The authors suggest that future work will involve exploring these 3D reconstruction methods for extraterrestrial exploration, specifically for autonomous robots or astronauts navigating the complex and unpredictable terrain of the Martian surface.

Summary updated

Abstract

<p indent="0mm">3D reconstruction is widely used in autonomous driving, resource exploration, etc. Wearable laser helmet sensor can achieve high-precision 3D reconstruction performance in areas without GNSS signals or limited access, due to its low-cost and miniaturized. However, during the data collection process, the wearable laser helmet sensor is prone to problems, such as severe jitter and rapid turns, which reduce the accuracy of trajectory estimation and 3D reconstruction results. To address these challenges, we first develop a point cloud motion compensation method based on intermediate reference points, which can reduce the motion distortion of point cloud data. Then, to improve the robustness, we propose an error model construction method based on local neighborhood feature consistency discrimination, which further improve the accuracy of sensor trajectory estimation and 3D reconstruction. Finally, we conducte a comprehensive comparative analysis of the method with the representative methods in public data sets, and designe a wearable laser helmet sensor to validate the effectiveness of the proposed method. Experimental results show that the proposed method achieves the best trajectory estimation and 3D reconstruction performance.
BibTeX
@inproceedings{gu2024solid,
  author = {Changjun Gu and Jiaxu Leng and ZongYi XU and XinBo GAO},
  title = {{A solid-state lidar-inertial 3D scene reconstruction method based on wearable laser helmet}},
  booktitle = {Scientia},
  year = {2024},
  doi = {10.1360/sst-2023-0362},
}
Figure