Collaborative Localization for Micro Aerial Vehicles

Sai Vemprala, Srikanth SaripalliIEEE Access 2021 · DOI

TL;DR

This paper introduces a vision-only framework for Micro Aerial Vehicles (MAVs) to collaboratively localize themselves using only onboard monocular cameras. By combining individual MAV pose estimates with relative pose information exchanged between MAVs, the system achieves more accurate and robust localization, especially in challenging environments where individual MAVs might struggle.

What Problem Does This Paper Address?

Micro Aerial Vehicles (MAVs) are increasingly used for various applications, but reliable localization is crucial for their autonomous operation. Monocular cameras, while compact and power-efficient, inherently suffer from scale ambiguity and can be prone to drift, especially in environments with sparse features or during rapid motion. Existing solutions often rely on additional sensors (like IMUs or GPS) or centralized systems, which can be costly, complex, or introduce single points of failure. This paper addresses the challenge of achieving robust and accurate 6-DoF localization for MAVs using only vision and inter-vehicle communication, aiming for a decentralized and resource-efficient solution.

Methods

The proposed Collaborative Localization (VCL) framework operates in a decentralized manner and consists of several key components:

  • Feature Detection and Matching: MAVs detect salient features (using AKAZE or KORAL) and match them to build a sparse 3D reconstruction of the environment, serving as a global map.
  • Intra-MAV Localization: Each MAV independently uses the global map to track features and estimate its own 6-DoF pose using a perspective-N-Point (PNP) algorithm, refined with a Kalman filter.
  • Inter-MAV Localization: When a MAV's individual localization is unreliable (e.g., due to lack of features), other MAVs can provide relative pose measurements. This involves:
  • Relative Pose Estimation: Using feature matches between two MAVs to estimate an essential matrix, from which relative rotation and translation are derived.
  • Scale Factor Estimation: Crucially, the scale ambiguity from relative pose estimation is resolved by matching the locally reconstructed features with the global map.
  • Guided Matching: A robust matching technique is employed to filter out incorrect correspondences between local and global map features.
  • Data Fusion: Individual (intra-MAV) and relative (inter-MAV) pose estimates are fused using a Covariance Intersection (CI) approach, which handles potentially unknown correlations between estimates.
  • Uncertainty Estimation: The framework estimates the uncertainty of the pose using the Jacobian of the reprojection error and propagates it through a Kalman filter.

Datasets

The paper evaluates the VCL algorithm using:

  • Simulated Data: Generated using Microsoft AirSim, a UAV simulator built on Unreal Engine, providing realistic rendering and ground truth. The simulations include urban environments and various flight trajectories.
  • Real-World Data: Collected from two Parrot Bebop 2 quadrotors, with onboard monocular camera footage.

No new benchmarks were introduced; existing datasets were used for evaluation.

Key Insights and Ideas

  • Vision-Only Collaboration: The framework demonstrates that robust localization can be achieved using only monocular cameras and inter-vehicle communication, eliminating the need for additional sensors like IMUs or GPS for collaborative localization.
  • Decentralized Approach: The VCL algorithm is designed to be decentralized, with each MAV performing its own localization and only communicating relative pose information when needed, reducing bandwidth requirements.
  • Inter-MAV as a Fallback: Inter-MAV localization acts as a powerful fallback mechanism when intra-MAV localization struggles due to feature scarcity, significantly improving robustness.
  • Scale Recovery is Key: Resolving the scale ambiguity in relative pose estimation by leveraging the global map is a critical step for achieving accurate metric poses.
  • Covariance Intersection for Fusion: Using CI for fusing estimates is effective even when correlations between sources are unknown, making it suitable for multi-MAV scenarios.

Key Results

  • Improved Accuracy: The VCL algorithm consistently shows improved pose estimation accuracy compared to intra-MAV localization alone, especially in challenging scenarios. For instance, in simulations, inter-MAV fusion reduced RMS error to less than 1% of the traversed distance.
  • Robustness to Motion: The system demonstrated accurate estimation of angles even during fast flights and pure rotations, where single-camera localization typically struggles.
  • Reduced Drift: In real-world experiments, inter-MAV fusion helped mitigate drift observed in intra-MAV localization when MAVs moved away from initial map features.
  • Scalability: The framework showed that increasing the number of contributing MAVs for inter-MAV fusion further improved localization accuracy.

Limitations and Future Work

  • Communication Delays: The current framework assumes no communication delays. Future work could investigate methods to handle these delays.
  • Real-time Deployment: While tested offline, the paper acknowledges the need for real-time deployment on onboard hardware.
  • More Complex Models: The Kalman filter currently uses a constant velocity model. Incorporating more complex vehicle dynamics or IMU measurements could further enhance performance.
  • SLAM Extension: Extending the framework to a full SLAM system, including map uncertainty, is a potential future direction.
  • Cooperative Assembly and Stereo Systems: The authors suggest applications in cooperative assembly and creating "decoupled aerial stereo" systems, which could benefit from the robust relative pose estimation capabilities.

Summary updated

BibTeX
@article{vemprala2021collaborative,
  author = {Sai Vemprala and Srikanth Saripalli},
  title = {{Collaborative Localization for Micro Aerial Vehicles}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3074537},
}
Figure