2D Simultaneous Localization And Mapping for Micro Air Vehicles

Adrien Angeli, David Filliat, Stéphane Doncieux, Jean-Arcady Meyer2006

TL;DR

This paper presents a vision-based Simultaneous Localization and Mapping (SLAM) system for Micro Air Vehicles (MAVs) operating at a constant altitude. The system uses visual odometry and an Extended Kalman Filter to simultaneously build a 2D map of visual landmarks and estimate the MAV's position, demonstrating robust loop-closure detection and improved map precision.

What Problem Does This Paper Address?

The increasing interest in Micro Air Vehicles (MAVs) has spurred research into autonomous navigation. A key challenge is enabling MAVs to navigate without prior knowledge of their environment, which requires them to simultaneously build a map and determine their own position within it – a problem known as Simultaneous Localization and Mapping (SLAM). While GPS is a common solution, this paper focuses on an onboard, vision-only approach, inspired by biological systems, to achieve autonomous flight. This is particularly relevant for MAVs where size, weight, and power constraints limit the use of traditional sensors like lidar or radar. The paper specifically targets 2D SLAM for MAVs flying at a constant altitude, which is a stepping stone towards more complex 3D navigation and behaviors like soaring.

Methods

The core of the proposed system is a vision-based SLAM approach using visual odometry coupled with an Extended Kalman Filter (EKF).

  • Visual Odometry: The system estimates the MAV's 2D displacement between consecutive images by tracking visual features. It combines the Harris corner detector with the SIFT descriptor for robust feature detection and matching, even with significant changes in viewpoint.
  • Extended Kalman Filter (EKF): The EKF is used to maintain and update the system's state, which includes the MAV's 2D position and orientation, as well as the 2D coordinates of detected landmarks in the map.
  • Prediction Step: The MAV's state is predicted based on the visual odometry measurements and the MAV's dynamic model.
  • Update Step: The predicted state is corrected using observations (detected landmarks) and their associated uncertainties.
  • Radial Distortion Estimation: To improve map precision, the system estimates the camera's radial distortion coefficient online. This is incorporated as an additional parameter in the EKF, allowing the system to correct for image warping and refine landmark positions.
  • Map Management: To handle the computational complexity of the EKF (O(N^2) where N is the number of landmarks), a landmark notation system is employed. This involves increasing a "recognition-score" for detected features and replacing older, less reliable landmarks with new ones, effectively managing the map size and enabling real-time processing.

Datasets

The paper evaluates the system using data collected from two types of MAVs:

  • A blimp (Figure 1)
  • The Paparazzi Twinstar MAV (Figure 2)

The experiments involve real-world flight data, with images sent to a ground-based PC for processing. No new benchmarks are introduced; rather, the system is tested on its ability to perform SLAM in real-world scenarios.

Key Insights and Ideas

  • Vision-only SLAM for MAVs is feasible: The paper demonstrates that a purely vision-based approach can achieve robust localization and mapping for MAVs, overcoming the limitations of traditional sensors.
  • SIFT for robust feature matching is crucial: Combining SIFT with feature tracking significantly improves the ability to match features across large baseline images, which is essential for loop-closure detection.
  • Online radial distortion estimation enhances map accuracy: Explicitly modeling and estimating camera distortion leads to more precise landmark positions and a more accurate map.
  • Loop closure dramatically reduces uncertainty: Detecting when the MAV revisits a previously mapped area significantly reduces both the MAV's position uncertainty and the uncertainty of the landmarks in the map.
  • voSLAM is a suitable paradigm for MAVs: The paper argues that a visual odometry-based SLAM (voSLAM) approach is more appropriate than Structure from Motion (SfM) for MAVs due to the need for a dynamic model and the desire for loop-closure detection.

Key Results

  • Successful Loop Closure: The system successfully detected loop closures in experiments with both a blimp and the Paparazzi MAV, leading to a significant reduction in position uncertainty.
  • Reduced Uncertainty: After loop closure, the MAV's position uncertainty was reduced to 5-10 pixels, and landmark uncertainties were around 1 pixel.
  • Accurate Radial Distortion Estimation: The radial distortion coefficient was estimated online and converged to a near-zero value, with its uncertainty decreasing over time, indicating accurate estimation.
  • Improved Trajectory: The EKF-filtered trajectory was shown to be more reliable and less prone to drift compared to raw visual odometry, especially in scenarios with significant movement. For example, Figure 5 shows the EKF trajectory (green) significantly outperforming the raw visual odometry (red) in a loop trajectory.

Limitations and Future Work

  • Data Association: The paper acknowledges that the data association problem (correctly matching features) is not entirely solved. False positive matches could still lead to Kalman filter divergence. The authors suggest using joint compatibility tests for matching groups of interest points as a potential solution.
  • Computational Complexity: The O(N^2) complexity of the EKF limits the number of landmarks that can be managed in real-time (around 150 features). Future work could involve combining global topological localization with precise metric mapping to handle larger environments.
  • 2D vs. 3D Navigation: The current system is limited to 2D navigation. Extending it to a 6-degree-of-freedom (3D) navigation system is a key area for future research, enabling more complex MAV behaviors.
  • Contextual Information: The authors suggest incorporating contextual information about landmarks into the map (e.g., identifying dangerous areas or recharging stations) to aid in trajectory planning.

Summary updated

Abstract

International audience
BibTeX
@misc{angeli2006d,
  author = {Adrien Angeli and David Filliat and Stéphane Doncieux and Jean-Arcady Meyer},
  title = {{2D Simultaneous Localization And Mapping for Micro Air Vehicles}},
  year = {2006},
}
Figure