GraphSLAM Improved by Removing Measurement Outliers

Ryun-Seok Kim, Hyukdoo Choi, Euntai KimJournal of Korean Institute of Intelligent Systems 2011 · DOI

Abstract

본 논문은 측정값의 우도를 기준으로 선택적인 측정값 적용을 통한 향상된 GraphSLAM을 제안하였다. GraphSLAM은 로봇의 이동 경로와 환경에 대한 지도를 전체 입력 데이터를 통해 추정한다. 그러나 잡음이 강한 환경에서 센서의 측정치가 부정확한 경우가 늘어나면, 전체 입력 데이터를 사용하는 GraphSLAM의 경우 정확성이 크게 떨어지게 된다. 그러므로 본 논문에서는 들어오는 센서의 측정값들을 선별하여 GraphSLAM에 적용하는 방법을 제안한다. 이 방법을 통해 잡음이 강한 환경에서 기존의 GraphSLAM보다 향상된 성능을 제공할 수 있다.

This paper presents the GraphSLAM improved by selecting the measurement with respect to their likelihoods. GraphSLAM estimates the robot's path and map by utilizing the entire history of input data. However, GraphSLAM's performance suffers a lot from severely noisy measurements. In this paper, we present GraphSLAM improved by the selective measurement method. Thus the presented GraphSLAM provides higher performance compared with the standard GraphSLAM.

Introduction

Simultaneous Localization and Mapping (SLAM) is a fundamental capability for autonomous robots, allowing them to build maps of unknown environments while tracking their own location. GraphSLAM is a popular formulation that treats the robot's trajectory and landmarks as nodes in a graph, with sensor measurements acting as edges or constraints between them.

GraphSLAM structure representing robot poses and landmarks with measurement constraints.

The primary weakness of standard GraphSLAM is its sensitivity to noise. In real-world scenarios, sensors often produce outliers—wildly incorrect measurements due to interference or hardware glitches. Because GraphSLAM attempts to satisfy all constraints globally, even a few outliers can significantly distort the resulting map and path estimate. This paper proposes a selective measurement approach to mitigate this issue.

Methodology: Selective GraphSLAM

The core contribution of this work is a framework that evaluates the statistical likelihood of each measurement before adding it to the graph. If a measurement is deemed an outlier, it is discarded, preventing it from corrupting the optimization process.

The selective measurement approach identifies and filters out outliers from the landmark measurement set.

The authors explore two filtering methods:

  1. Empirical Approach: This method calculates the mean and variance of a landmark's position based on historical data. New measurements that deviate significantly from this average are flagged as outliers.
  2. Theoretical Approach: A more robust method that utilizes the Jacobian of the measurement model and the system's covariance matrix. By accounting for both sensor noise and the robot's own localization uncertainty, the system becomes more skeptical of new data when its own position estimate is shaky.

In both cases, a binary weighting system is used: measurements with a likelihood above a threshold $\tau$ are kept (weight 1), while others are discarded (weight 0).

Experimental Results

The proposed methods were tested using MATLAB simulations in both low-noise and high-noise (5x standard noise) scenarios.

In low-noise environments, the theoretical selective method provided a modest 10% improvement in localization accuracy. However, in high-noise environments, the impact was profound:

  • Standard GraphSLAM: Localization error reached approximately 324 cm.
  • Empirical Selective Method: Improved localization and mapping by 12-13%.
  • Theoretical Selective Method: Reduced localization error by 43% (down to ~186 cm) and mapping error by 47% compared to the standard approach.

These results demonstrate that the theoretical approach effectively prevents error compounding by making the robot "self-aware" of its own estimation uncertainty.

Conclusion and Future Work

Selective measurement significantly improves GraphSLAM's robustness in noisy environments. By filtering outliers based on theoretical likelihood, the system achieves nearly double the accuracy of standard methods under high interference. Future research will focus on developing adaptive thresholds for outlier detection and validating the algorithm on physical hardware to address non-Gaussian real-world noise.

Summary updated

Abstract

본 논문은 측정값의 우도를 기준으로 선택적인 측정값 적용을 통한 향상된 GraphSLAM을 제안하였다. GraphSLAM은 로봇의 이동 경로와 환경에 대한 지도를 전체 입력 데이터를 통해 추정한다. 그러나 잡음이 강한 환경에서 센서의 측정치가 부정확한 경우가 늘어나면, 전체 입력 데이터를 사용하는 GraphSLAM의 경우 정확성이 크게 떨어지게 된다. 그러므로 본 논문에서는 들어오는 센서의 측정값들을 선별하여 GraphSLAM에 적용하는 방법을 제안한다. 이 방법을 통해 잡음이 강한 환경에서 기존의 GraphSLAM보다 향상된 성능을 제공할 수 있다. This paper presents the GraphSLAM improved by selecting the measurement with respect to their likelihoods. GraphSLAM estimates the robot's path and map by utilizing the entire history of input data. However, GraphSLAM's performance suffers a lot from severely noisy measurements. In this paper, we present GraphSLAM improved by the selective measurement method. Thus the presented GraphSLAM provides higher performance compared with the standard GraphSLAM.
BibTeX
@article{kim2011graphslam,
  author = {Ryun-Seok Kim and Hyukdoo Choi and Euntai Kim},
  title = {{GraphSLAM Improved by Removing Measurement Outliers}},
  journal = {Journal of Korean Institute of Intelligent Systems},
  year = {2011},
  doi = {10.5391/jkiis.2011.21.4.493},
}
Figure