A Fast Robot Identification and Mapping Algorithm Based on Kinect Sensor

Liang Zhang, Peiyi Shen, Guangming Zhu, Wei Wei, Houbing Song2015 · DOI

TL;DR

This paper tackles the challenge of real-time, accurate Simultaneous Localization and Mapping (SLAM) for robots using Kinect depth cameras. They propose an improved RGB-D SLAM algorithm that significantly enhances speed and precision by replacing traditional feature detection, matching, and motion estimation methods with more efficient alternatives like ORB, FLANN KNN, and an improved RANSAC/GICP pipeline.

What Problem Does This Paper Address?

Autonomous robots need to understand their environment to navigate and perform tasks. Simultaneous Localization and Mapping (SLAM) is a crucial capability that allows a robot to build a map of its surroundings while simultaneously tracking its own position within that map. RGB-D SLAM, which uses depth information from sensors like the Kinect, has emerged as a powerful approach. However, existing RGB-D SLAM algorithms often suffer from two main drawbacks: insufficient speed for real-time applications and unacceptable accuracy. These limitations hinder their practical deployment in dynamic or complex environments.

Methods

The authors propose an improved RGB-D SLAM algorithm that enhances both the "front-end" (feature detection, matching, and motion estimation) and "back-end" (pose graph optimization) components.

Key improvements include:

  • Feature Detection and Descriptor Extraction: Replaced SIFT and SURF with the ORB (Oriented FAST and Rotated BRIEF) method. ORB is known for its speed and robustness, and the paper also highlights its ability to filter illegal depth information, which can reduce descriptor count and computation.
  • Feature Matching: Replaced the brute-force method with a bidirectional FLANN (Fast Library for Approximate Nearest Neighbors) KNN (k-Nearest Neighbor) algorithm. FLANN is an efficient approximate nearest neighbor search library, and the bidirectional approach helps to eliminate false matches by considering matches from both forward and reverse directions. Homography transformation is also used to further optimize matching results.
  • Motion Estimation: Introduced an improved RANSAC (RANdom SAmple Consensus) method. This iterative approach refines the motion transformation model by re-evaluating inliers, leading to more accurate results.
  • Motion Transformation Optimization: Replaced the standard ICP (Iterative Closest Points) with GICP (Generalized ICP), which uses a probabilistic model and surface information for more accurate point cloud registration. The paper also discusses how GICP can be degenerated to ICP when necessary.
  • Closed-Loop Detection: The paper mentions the use of a DCS (Dynamic Covariance Scaling) algorithm within the G2O framework for closed-loop detection, with an analytical solution derived for a weighting factor.

Datasets

The paper primarily uses the rgbd_dataset_freiburg1_360 dataset, which contains 754 frames of RGB and Depth images. They also mention testing on the Freiburg dataset for specific scenarios and using a real Dr Robot X80 environment with 475 RGB-D images for actual deployment validation.

Key Insights and Ideas

  • ORB's Efficiency: The ORB feature detector and descriptor extraction significantly outperform SIFT and SURF in terms of speed, making real-time processing more feasible.
  • Bidirectional Matching is Crucial: Simply using FLANN KNN isn't enough; the bidirectional approach is key to eliminating false matches and improving the robustness of feature matching.
  • GICP for Precision: While ICP is standard, GICP's ability to leverage surface information and probabilistic modeling leads to more accurate point cloud registration, a critical step for precise mapping.
  • Iterative Refinement in RANSAC: The improved RANSAC's iterative nature, where inliers are re-evaluated, is vital for obtaining a more reliable motion transformation model, especially in the presence of outliers.
  • Addressing Failures: The proposed methods, particularly the improved back-end, demonstrate the ability to handle situations where original methods might fail (e.g., SURF_0 in motion transformation estimation).

Key Results

  • Speed Improvement: The proposed ORB-based feature detection and FLANN KNN matching significantly reduce computation time. For example, the ORB method for feature detection, descriptor extraction, and improved bidirectional FLANN KNN/homography transformation took only 37 ms in one comparison.
  • Accuracy Enhancement: The improved RANSAC and GICP methods lead to lower deviation in motion transformation estimation. For instance, in Table 5, the improved methods consistently show lower deviation values compared to the original methods.
  • Reduced Errors: The improved front-end algorithm shows reduced translational and rotational errors, particularly for methods like SURF_0 and ORB_10000, compared to the original algorithm.
  • Robustness: The improved algorithm demonstrates better performance in actual environment tests on the Dr Robot X80, with better splicing results and a more accurate robot trajectory.

Limitations and Future Work

  • Limitations:
    • The paper notes that the translational error staying around 0.05 m and rotational error being 0.05° in some experiments might depend on the inherent error of the Kinect camera itself.
  • GICP can fail if the number of matching inliers is too few (less than 20 pairs), though they propose three methods to address this.
  • While the improved algorithm is faster and more accurate, the improved motion transformation optimization might spend 1.5–3 times longer than the original method, though the error reduction is substantial.
  • Future Work: The paper concludes by stating that the improved algorithm solves low speed and low accuracy problems, satisfying high speed and high precision requirements. This makes it suitable for IoT systems constructed by robots, enhancing automation and real-time robustness. Further applications in robot identification and mapping within IoT systems are implied.

Summary updated

BibTeX
@misc{zhang2015fast,
  author = {Liang Zhang and Peiyi Shen and Guangming Zhu and Wei Wei and Houbing Song},
  title = {{A Fast Robot Identification and Mapping Algorithm Based on Kinect Sensor}},
  year = {2015},
  doi = {10.3390/s150819937},
}
Figure