TL;DR
This paper presents a robust vision-based system for quadrotor UAVs to accurately position and land on moving targets, even under challenging conditions like shadows and partial occlusion. By combining novel image processing, Kalman filtering for motion estimation, and a Model Predictive Control (MPC) strategy, the system achieves high landing accuracy and reliable tracking, paving the way for more autonomous aerial operations.
What Problem Does This Paper Address?
Autonomous landing for quadrotor UAVs, especially on moving platforms, is a critical but challenging task. Existing methods often struggle with:
- Shadow Occlusion: Shadows cast by the UAV or its surroundings can obscure target markers, hindering recognition.
- Field of View Limitations: As the UAV descends, the target can become partially or fully out of view, leading to loss of tracking.
- Dynamic Targets: Landing on a moving platform requires precise tracking of its position and velocity, which is difficult in real-world, unpredictable environments.
- Environmental Factors: Barometric and GPS altitude readings can be unreliable, especially at low altitudes or in challenging weather, making accurate height estimation crucial for safe landing.
- Controller Performance: Standard controllers may not offer sufficient accuracy or responsiveness for precise landing on dynamic targets.
This paper aims to overcome these limitations by developing a comprehensive system that can reliably recognize, track, and land on moving targets in various conditions.
Methods
The proposed system integrates several key components:
- Novel Image Preprocessing: To combat shadow occlusion, the authors employ a finite adaptive histogram equalization technique in the HSV color space. This method specifically targets the saturation channel, which is most affected by shadows, to enhance marker visibility without excessive computational cost.
- Binocular Vision-Based Target Recognition and Localization:
- Apriltag Markers: A nested design of Apriltags of different sizes is used to ensure at least one marker is always visible within the camera's field of view, regardless of UAV altitude.
- Kalman Filter (KF/EKF): A Kalman filter-based approach is used for target motion state estimation, allowing the UAV to predict and track the target's position even during brief periods of occlusion. The Extended Kalman Filter (EKF) is found to perform better under dynamic conditions.
- Binocular Vision for Depth: A stereo vision-based depth camera is used to estimate the relative height between the UAV and the target. This is crucial for accurate landing, especially when barometric sensors are unreliable. The system compensates for UAV attitude to improve depth measurement accuracy.
- Feedforward Model Predictive Control (MPC): To address controller performance issues, a feedforward MPC algorithm is integrated. This controller focuses on minimizing relative position and velocity errors between the UAV and the target, enabling robust tracking of both stationary and moving targets.
- Airspace Division Strategy: A novel mobile tracking and landing control strategy based on dividing the airspace around the target is proposed. This strategy dynamically adjusts the UAV's descent speed based on its relative position to the target, enhancing safety and success rates.
Datasets
Not discussed in the paper. The experiments rely on real-world flight tests and simulations using custom setups and parameters.
Key Insights and Ideas
- HSV Saturation for Shadows: Focusing image processing on the saturation channel in HSV space is a clever way to enhance features obscured by shadows without needing complex, computationally expensive algorithms.
- Nested Apriltags for Robustness: Using multiple sizes of Apriltags is a simple yet effective solution to the problem of target features disappearing from view as the UAV descends.
- Binocular Vision for Height is Crucial: Relying solely on barometers or GPS for altitude is insufficient for precise landing. Stereo vision, combined with attitude compensation, provides a more reliable height estimate.
- MPC for Dynamic Precision: MPC's ability to predict future states and optimize control inputs makes it well-suited for the complex task of landing on a moving target, offering better dynamic performance than traditional PID controllers in some scenarios.
- Airspace Division for Safety: Segmenting the landing approach into different airspaces allows for a more controlled and adaptive descent, reducing the risk of errors and improving safety.
Key Results
- Target Recognition: 100% success rate.
- Positioning Accuracy:
- Close range: x and y-axis errors not exceeding 0.01 m.
- Medium range: x-axis relative error not exceeding 0.05 m, y-axis error not exceeding 0.03 m.
- Long range: relative errors for both axes not exceeding 0.05 m.
- Tracking Accuracy: EKF outperforms KF in dynamic conditions, offering better estimation and faster tracking.
- Landing Accuracy: Within 0.1 m.
- Mobile Energy Supply Mission: The proposed method successfully accomplished this mission in experiments.
- MPC vs. PID: While MPC might show slightly larger errors in the final landing phase due to its strategy, it demonstrates superior control dynamics and faster initial target acquisition compared to PID.
Limitations and Future Work
- MPC Landing Error: The paper notes that the MPC controller can result in a slightly larger landing error than PID due to its strategy of abruptly powering down, which can cause rebound.
- Dissimilar PID Parameters: Differences in PID parameters for the x and y axes are cited as a reason for slight variations in landing errors.
- Computational Load: While efforts were made to optimize processing, real-time performance on embedded systems remains a consideration.
- Future Work: The paper suggests further research into improving the MPC controller's precision in the final landing phase and potentially exploring more advanced control strategies. The successful demonstration of the mobile energy supply mission hints at broader applications.
Summary updated