[Paper Review] Localising Faster: Efficient and precise lidar-based robot localisation in large-scale environments

Li Sun, Daniel Adolfsson+4 more|arXiv (Cornell University)|Mar 4, 2020

Robotics and Sensor-Based Localization30 references4 citations

TL;DR

This paper proposes a hybrid lidar-based localisation method that accelerates Monte Carlo Localisation (MCL) by seeding it with a deep Gaussian process model for 6-DOF pose estimation. By integrating learning-based inference with geometric verification via importance sampling, the system achieves sub-second average localisation (1.94 s) with 0.75 m precision in large-scale environments, reducing MCL runtime by 99.5% compared to baseline.

ABSTRACT

This paper proposes a novel approach for global localisation of mobile robots in large-scale environments. Our method leverages learning-based localisation and filtering-based localisation, to localise the robot efficiently and precisely through seeding Monte Carlo Localisation (MCL) with a deep-learned distribution. In particular, a fast localisation system rapidly estimates the 6-DOF pose through a deep-probabilistic model (Gaussian Process Regression with a deep kernel), then a precise recursive estimator refines the estimated robot pose according to the geometric alignment. More importantly, the Gaussian method (i.e. deep probabilistic localisation) and non-Gaussian method (i.e. MCL) can be integrated naturally via importance sampling. Consequently, the two systems can be integrated seamlessly and mutually benefit from each other. To verify the proposed framework, we provide a case study in large-scale localisation with a 3D lidar sensor. Our experiments on the Michigan NCLT long-term dataset show that the proposed method is able to localise the robot in 1.94 s on average (median of 0.8 s) with precision 0.75~m in a large-scale environment of approximately 0.5 km2.

Motivation & Objective

• To address the challenge of efficient and precise global localisation in large-scale, GPS-denied environments such as indoor industrial sites or underground mines.
• To overcome the slow convergence of Monte Carlo Localisation (MCL) when initialised with uniform particle distributions in large maps.
• To integrate the speed of deep learning-based pose estimation with the geometric accuracy of filtering-based methods like MCL.
• To enable real-time, precise localisation in large-scale environments (e.g., 0.5 km²) using only 3D lidar data.
• To develop a seamless fusion mechanism between Gaussian (GP) and non-Gaussian (MCL) probabilistic models via importance sampling.

Proposed method

• A deep probabilistic model based on Gaussian Process Regression with a deep kernel is trained to estimate the 6-DOF pose distribution from superimposed lidar scans.
• The estimated Gaussian process distribution is used to seed MCL particles, replacing uniform or heuristic initialisations.
• Importance sampling is employed to integrate the GP’s predictive distribution with MCL, enabling mutual benefit between learning-based and filtering-based systems.
• A sparse Normalised Dense Transformation (NDT) model is used for efficient measurement likelihood computation in MCL.
• Particles are resampled and updated recursively using MCL, with dynamic particle count (500–1000) and adaptive resampling to maintain efficiency.
• The system uses a voxel grid with 0.2 m resolution to limit particle count and improve computational tractability.

Experimental results

Research questions

• RQ1 Can a deep Gaussian process model provide a reliable, informative prior for MCL in large-scale lidar environments?
• RQ2 How can the non-conjugate combination of a Gaussian process (GP) and a non-Gaussian particle filter (MCL) be effectively integrated?
• RQ3 Does seeding MCL with a learned GP distribution significantly reduce localisation time compared to uniform or fixed initialisation?
• RQ4 To what extent does the uncertainty estimate from the GP improve robustness against false positives in long-term, dynamic environments?
• RQ5 Can the hybrid system achieve centimetre-level precision while maintaining sub-second average localisation time in large-scale maps?

Key findings

• The hybrid method achieves an average localisation time of 1.94 seconds (median 0.8 seconds) on the Michigan NCLT dataset, a 99.5% reduction compared to the MCL baseline.
• The system localises with a precision of 0.75 m in a large-scale environment of approximately 0.5 km², meeting real-world requirements for autonomous navigation.
• The success rate of the hybrid method was highest in early months (e.g., January–April), decreasing in November and December due to increased novel trajectories and landmarks.
• Using GP-based uncertainty for particle seeding reduced localisation time by 99.5% compared to the MCL baseline, which had a 54% success rate and averaged 154.3 seconds per attempt.
• The method with GP-based covariance outperformed fixed-covariance seeding, showing that uncertainty-aware particle sampling significantly improves convergence speed and reliability.
• The integration of deep learning and filtering via importance sampling enables seamless, mutually beneficial operation between fast inference and geometric verification.