TL;DR
This paper presents a vision-based autonomous landing system for quadcopters that can track and land on moving targets. By combining a pure pursuit guidance law with a novel log-polynomial closing velocity controller, the system achieves faster and more accurate landings compared to traditional methods, even in challenging scenarios like target occlusion.
What Problem Does This Paper Address?
Autonomous landing for quadcopters is a critical but challenging task, especially when the landing target is moving. Traditional methods often rely on precise GPS information, which can be inaccurate, or require communication between the UAV and the target. Vision-based approaches are promising but face difficulties with dynamic targets, especially when the target's movement is unpredictable or the UAV's field of view is limited. This paper addresses the challenge of achieving fast, accurate, and robust autonomous landings on moving targets using only onboard vision, without relying on external communication or precise prior knowledge of the target's trajectory.
Methods
The core of the proposed system is a vision-based guidance and control architecture.
- Vision System: The system uses a combination of color segmentation (HSV and YCrCb color spaces) and AprilTags to detect the landing pad. This dual approach enhances robustness across varying altitudes and lighting conditions. A Kalman Filter is employed for target state estimation, providing smoothed position and velocity estimates.
- Guidance Law: The paper utilizes a Pure Pursuit (PP) guidance law, which is known for its simplicity and effectiveness in keeping the target within the vehicle's field of view.
- Closing Velocity Controller: A novel log-polynomial function is introduced to control the quadcopter's closing velocity (the speed at which it approaches the target). This controller ensures that the closing velocity asymptotically approaches zero for a safe landing. It also provides a desirable velocity profile: a slower decay during most of the flight for consistent tracking, followed by a faster decay towards the end to quickly minimize closing velocity. This controller is applied independently to both lateral/longitudinal and vertical movements.
Datasets
Not discussed in the paper. The paper relies on simulations and real-world experiments.
Key Insights and Ideas
- Dual Detection Strategy: Combining color segmentation with AprilTags provides a robust landing pad detection mechanism that works well from both high and low altitudes.
- Log-Polynomial Velocity Control: This approach offers fine-grained control over the UAV's velocity profile, enabling time-efficient landings while maintaining target visibility. The ability to shape the velocity decay is crucial for balancing speed and precision.
- Vision-Only Approach: The system's reliance solely on onboard vision, without external communication or prior target trajectory knowledge, makes it more practical and adaptable for real-world scenarios.
- Robustness to Occlusion: The use of a Kalman Filter for target state estimation allows the system to maintain a predicted trajectory even during temporary target occlusions, significantly improving robustness.
Key Results
The paper reports on both simulations and extensive real-world experiments.
- Experimental Success Rate: Out of 27 field experiment runs on a moving target, 22 successful landings were achieved.
- Landing Accuracy: The maximum error magnitude for successful landings was recorded at 35 cm from the landing target center. In 13 of these runs, the landing accuracy was within ±20 cm.
- Simulation Performance: Simulations demonstrated successful landings on targets moving in straight lines (up to 3 m/s) and circular trajectories. The system also showed robustness to noisy target information.
- Time Efficiency: The "track and descend" approach, enabled by the log-polynomial controller, leads to more time-efficient landings compared to traditional track-then-vertical-land methods. The fastest landing achieved was 41 seconds.
- Failure Analysis: Failures were primarily attributed to target loss due to occlusion or estimation errors at low altitudes, highlighting areas for future improvement.
Limitations and Future Work
- Target Speed: Current experiments were limited to lower target speeds (up to 0.6 m/s in field tests, 3 m/s in simulations). Testing at higher speeds is a future goal.
- Environmental Disturbances: While tested in mild wind, the robustness to significant wind disturbances needs further investigation.
- Parameter Selection: Developing an automated method for selecting log-polynomial parameters for high-speed landing and in the presence of estimation errors is a key area for future work.
- 3D Motion and Complex Environments: Extending the approach to targets with 3D motion (e.g., on ships) and environments with significant wind disturbances is suggested.
- Sensor Fusion: Exploring different sensor fusion approaches for landing and comparing them with the current vision-based method is also proposed.
- Hybrid Approach for Target Loss: For scenarios where the target is lost for an extended period, a hybrid approach involving a search strategy with increased UAV altitude is planned.
Summary updated