TL;DR
This paper presents a robust, on-board vision-based system for micro aerial vehicles (MAVs) to safely navigate dynamic environments with moving obstacles. By combining efficient depth-image-based obstacle detection and tracking with a chance-constrained model predictive controller (CC-MPC), the MAV can generate safe, dynamically feasible trajectories while accounting for sensing and state estimation uncertainties.
What Problem Does This Paper Address?
Autonomous navigation for Micro Aerial Vehicles (MAVs) is crucial for applications like search and rescue, inspection, and cinematography. While MAVs can navigate static environments or controlled dynamic ones with external systems, safely and robustly avoiding moving obstacles in unknown, dynamic environments using only on-board sensors remains a significant challenge. Existing methods often struggle with real-time performance, handling uncertainties in obstacle detection and MAV state estimation, or accounting for the MAV's own dynamics. This paper aims to bridge this gap by providing an integrated, on-board solution that addresses these challenges.
Methods
The core of the proposed system lies in its integrated approach:
- Obstacle Detection and Tracking: The system leverages depth images from a stereo camera to detect and track obstacles. It extends previous work by extending obstacle representation from 2D to 3D ellipsoids. An "U-depth map" is computed from depth image histograms to efficiently identify potential obstacles. This is followed by a Kalman filter to track the position, velocity, and size of detected obstacles, along with their uncertainties.
- Robust Collision Avoidance: The detected obstacle information, including uncertainties, is fed into a Chance-Constrained Model Predictive Controller (CC-MPC). This controller formulates constraints probabilistically, ensuring that the collision probability between the MAV and each obstacle remains below a predefined threshold (δ). This probabilistic approach is key to achieving robustness against sensing and state estimation uncertainties.
- MAV Dynamics and Constraints: The CC-MPC explicitly considers the MAV's nonlinear dynamics and incorporates several cost functions: goal navigation, control input minimization, and an obstacle potential field for safety. Crucially, it also includes Field of View (FOV) constraints to ensure the MAV stays within its sensing capabilities.
- Uncertainty Propagation: To efficiently handle the propagation of MAV state uncertainty through the prediction horizon, an Extended Kalman Filter (EKF) based update is employed, which is computationally less intensive than full nonlinear propagation.
Datasets
The paper does not explicitly mention training datasets. Instead, it focuses on real-world experimental evaluation.
- Evaluation: Experiments were conducted in a lab setting with two walking human obstacles and in a long corridor with static and moving obstacles. Ground truth data for evaluation was collected using an external motion capture system (OptiTrack).
Key Insights and Ideas
- Probabilistic Robustness is Key: The most significant insight is the use of chance constraints within the MPC framework. This moves beyond deterministic collision avoidance to a probabilistic guarantee, making the system inherently more robust to the inherent uncertainties in real-world sensing and state estimation.
- Integrated System Design: The paper demonstrates the effectiveness of integrating fast, on-board obstacle detection/tracking with a robust motion planner, rather than relying on separate, potentially disconnected modules.
- Efficient Uncertainty Handling: The use of an EKF for uncertainty propagation, while propagating based on the previous loop's trajectory, is a practical engineering choice to balance accuracy and real-time performance for the MPC.
- 3D Obstacle Representation: Extending obstacle representation to 3D ellipsoids from 2D bounding boxes provides a more accurate model for collision checking, especially in dynamic scenarios.
Key Results
- Effective Collision Avoidance: Flight tests demonstrated successful avoidance of moving human obstacles in various scenarios, including a confined lab space and a long corridor.
- Quantitative Performance:
- Obstacle detection and tracking achieved an average position estimation error of around 0.3 m and a velocity estimation error of up to 0.5 m/s, highlighting the need for uncertainty handling.
- The system achieved a minimum safe separation of 0.4 m from obstacles in experiments.
- On-board runtimes were impressive: obstacle detection and tracking averaged 8 ms (allowing for 60 Hz operation), and the MPC ran at 16 ms (allowing for ~60 Hz operation), demonstrating real-time capability on a limited computational platform.
- MAV speeds up to 1.6 m/s in the lab and 2.4 m/s in the corridor were achieved while maintaining safety.
Limitations and Future Work
- Obstacle Size Uncertainty: The paper states that obstacle size uncertainty is not considered in collision avoidance. This could be a limitation in scenarios with highly variable obstacle shapes or sizes.
- Limited FOV: While FOV constraints are included, the system's performance is inherently limited by the camera's field of view.
- Computational Power: Although the system is designed for on-board computation, further improvements in computational efficiency could allow for higher prediction horizons or more complex environments.
- Future Work: The authors suggest exploring more advanced uncertainty propagation methods and potentially incorporating more sophisticated obstacle models. Further testing in more complex and cluttered environments with a higher density of moving obstacles would also be valuable.
Summary updated