Cooperative LiDAR SLAM for multi-vehicles based on edge computing

Mingyue Cui, Shipeng Zhong, Siyao Liu, Boyang Li, Chenghao Wu, Kai HuangJournal of Image and Graphics 2021 · DOI

Abstract LiDAR-based Simultaneous Localization and Mapping (SLAM) is a cornerstone of intelligent robotics, enabling autonomous vehicles to navigate through 3D environmental modeling. However, individual vehicles often struggle with long mapping cycles and high computational demands. This research proposes a multi-vehicle collaborative mapping method based on edge computing. By utilizing a threshold-based offloading function and modeling task scheduling as a potential game, the system achieves load balancing and high map reusability. Experimental results demonstrate that this cooperative approach significantly enhances positioning accuracy—improving lateral and longitudinal precision by 6.0x and 3.9x, respectively—while reducing local resource requirements and latency.

Introduction

As autonomous driving technology advances, the need for high-precision, real-time environmental perception becomes critical. LiDAR SLAM provides the necessary 3D modeling for navigation, but the sheer volume of data often exceeds the processing capabilities of onboard units.

Autonomous driving systems require complex environmental perception and real-time processing.

This paper introduces a framework where multiple vehicles collaborate with an edge server to distribute the computational load. This not only speeds up the mapping process but also allows vehicles to benefit from a shared, more comprehensive global map.

The Challenge of Single-Vehicle SLAM

Traditional SLAM systems operating on a single vehicle face two primary bottlenecks:

  1. Computational Intensity: Real-time processing of massive point cloud data requires significant power, leading to high latency and potential safety risks.
  2. Accumulated Error: A single vehicle's perspective is limited. Over long distances, odometry errors accumulate, leading to "drift" in the generated map.

While multi-vehicle systems can mitigate these issues, they introduce the "offloading problem": determining which tasks should stay on the vehicle and which should be sent to the edge server to avoid network congestion.

Methodology

1. Game Theory for Task Offloading

To optimize resource allocation, the researchers modeled the offloading decision as a Potential Game. In this model, each vehicle acts as a player seeking to minimize its own latency and maximize signal quality. The authors proved that this system possesses a Nash Equilibrium, a stable state where no vehicle can improve its performance by changing its offloading strategy alone. To reach this state quickly, they implemented $\alpha$-Nash best response dynamics, an iterative algorithm that accelerates convergence even as the number of vehicles scales up.

2. Coarse-to-Fine Map Merging

Merging maps from different vehicles requires high precision to avoid ghosting or misalignment. The system uses a two-step process:

  • Coarse Matching: Employs Fast Point Feature Histograms (FPFH) and RANSAC to establish an initial alignment between local maps.
  • Fine Matching: Refines the alignment using the Iterative Closest Point (ICP) algorithm.
  • Relative Confidence Metric: A novel weighting system that prioritizes map segments based on their reliability (calculated from odometry distance and overlap size), ensuring the global map remains accurate over time.

Experimental Validation

The framework was tested through both large-scale simulations (up to 50 vehicles) and real-world trials. The physical experiment involved four electric patrol vehicles equipped with Velodyne LiDAR and NVIDIA Tegra X2 units, navigating a 1.49 km campus route. High-precision RTK-DGPS was used as the ground truth for performance evaluation.

Key Results and Discussion

The cooperative approach yielded substantial improvements over standalone SLAM:

  • Precision Gains: Lateral positioning accuracy improved by 6.0x, while longitudinal accuracy improved by 3.9x.
  • Drift Correction: By merging data from multiple perspectives, the system effectively "closed the loop" on mapping errors, resulting in trajectories that closely matched the RTK-DGPS ground truth.
  • Scalability: The potential game-based scheduling proved efficient, maintaining low latency even as vehicle density increased.

Conclusion

By leveraging edge computing and game theory, this research provides a robust solution for multi-vehicle LiDAR SLAM. The method effectively balances the computational load across the network while delivering high-precision localization that rivals expensive differential GPS systems. Future work will focus on incorporating energy consumption constraints and improving robustness in highly dynamic environments with moving obstacles.

Summary updated

Abstract

目的激光雷达实时定位与建图(simultaneous localization and mapping,SLAM)是智能机器人领域的重要组成部分,通过对周边环境的3维建模,可以实现无人驾驶车辆的自主定位和精准导航。针对目前单个车辆激光雷达建图周期长、算力需求大的现状,提出了基于边缘计算的多车协同建图方法,能够有效地负载均衡,在保证单个车辆精准定位的同时,增加多个车辆之间的地图重用性。方法构建基于阈值的卸载函数,论证边缘计算下的多车卸载决策属于势博弈问题,设计实现基于边缘计算的势博弈卸载算法,在模型具有纳什均衡的基础上实现任务调度,引入<inline-formula><tex-math id="M1">$\alpha $</tex-math></inline-formula>-Nash最佳响应动态加速算法收敛,并采用由粗到细的点云匹配方法提高地图匹配性能,实现车辆的精准定位。最后,基于地图的相对可信度,高效地合并基站覆盖范围内的多个车辆的建图数据。结果实验表明,基于博弈论的调度方法在保证定位可靠性的前提下,能够有效地实现多车协同SLAM,且多车协同的定位与建图结果与使用载波相位差分技术(real-time kinematic,RTK)的高精度差分全球定位系统(differential global positioning system,DGPS)结果足够接近,相比于单车建图而言,横向定位和纵向定位的平均精度分别提高了6.0倍和3.9倍。结论本文方法解决了基于边缘计算的多车协同激光雷达SLAM问题,借助边缘服务器的计算资源,无人驾驶车辆可以有效地减少本地资源需求和定位延迟。该方法通过各个车辆之间的资源博弈,最终实现纳什均衡。实现基于边缘计算的激光雷达定位服务,且高效地完成多车之间的地图合并,仿真和真实环境中的实验表明了方法的有效性。
BibTeX
@article{cui2021cooperative,
  author = {Mingyue Cui and Shipeng Zhong and Siyao Liu and Boyang Li and Chenghao Wu and Kai Huang},
  title = {{Cooperative LiDAR SLAM for multi-vehicles based on edge computing}},
  journal = {Journal of Image and Graphics},
  year = {2021},
  doi = {10.11834/jig.200441},
}
Figure