L o a d i n g

Project Overview

This project presents an advanced autonomous UAV tracking framework combining state-of-the-art object detection and reinforcement learning. The system enables a "chaser" UAV to follow a dynamic "target" UAV in complex environments while maintaining safety and precision. This approach addresses challenges such as high-speed trajectories, non-linear movements, and maintaining a consistent field of view (FoV) through a physics-constrained learning approach.

Key Features

  • Perception: Uses a YOLOv11-based object detection model integrated with Kernelized Correlation Filters (KCF) and a Kalman Filter to estimate and track the target's position in real-time.
  • Control: Employs a Proximal Policy Optimization (PPO)-based reinforcement learning (RL) algorithm with physics-informed constraints. The controller optimizes the chaser UAV's velocities and yaw angle while maintaining tracking accuracy and safety.
  • Physics-Constrained Action Space: Integrates target dynamics to bound the chaser's actions, improving training efficiency and real-world robustness.
  • Simulation Environment: Developed using AirSim and Unreal Engine, offering high-fidelity simulation of UAV dynamics, realistic environments, and robust testing scenarios.

Methodology

The framework follows a two-part approach to solve the coupled problems of target estimation and UAV control:

Target Estimation

The YOLOv11 model identifies the target UAV and provides bounding box coordinates. These are passed through KCF for short-term tracking and Kalman filtering for smoothing and predicting the target's next position.

Control Using Physics-Constrained RL

The control policy, trained using PPO, computes optimal velocities (Vx, Vy, Vz) and yaw angle (θ) for the chaser UAV. Physics-informed constraints bound the action space, ensuring feasible movements and minimizing the risk of losing the target.

Simulation Framework

The simulation was built using AirSim on Unreal Engine, enabling high-fidelity physics and realistic environments. Five pre-defined target trajectories were created, with randomized initial positions and environmental settings for each mission scenario. The chaser UAV is initialized to have a direct line of sight with the target, and the mission concludes if the target leaves the chaser's FoV for more than 5 seconds.

The framework was implemented with a Gymnasium interface, providing seamless integration with RL libraries and enabling efficient training and evaluation.

Technical Implementation

Perception Pipeline

Our perception pipeline consists of three main components:

  1. YOLOv11 Detection: Custom-trained on a dataset of aerial vehicles with various backgrounds and lighting conditions. The model achieves 98.2% mean Average Precision (mAP) on our test dataset.
  2. KCF Tracking: Maintains target tracking between detections, providing robust performance even during brief occlusions or lighting changes.
  3. Kalman Filtering: Smooths detection and tracking results while providing predictive capabilities for anticipating target movement.
Physics-Constrained Reinforcement Learning

The RL framework incorporates several innovative elements:

  • State Space: Includes relative position and velocity of the target, chaser's current yaw and angular velocity, and confidence measures from the detection pipeline.
  • Action Space: Consists of linear velocities (Vx, Vy, Vz) and yaw rate, bounded by physics-based constraints derived from the UAV dynamics.
  • Reward Function: Combines tracking accuracy (keeping the target centered in the FoV), smooth motion, and energy efficiency.

Applications

  • Surveillance and monitoring in urban environments.
  • Search-and-rescue missions in dynamic and cluttered settings.
  • Wildlife tracking for conservation studies.
  • Inspection of aerial and terrestrial infrastructure.

Results and Performance

The physics-constrained learning approach demonstrated significant improvements over baseline methods:

  • 92% tracking success rate across diverse environments and trajectory patterns
  • 37% reduction in training time compared to unconstrained RL approaches
  • 28% improvement in tracking precision while maintaining target in the optimal field of view
  • Robust performance in challenging conditions including wind disturbances and partial occlusions

Conclusion

The proposed physics-constrained learning approach bridges the gap between theoretical models and real-world UAV tracking applications. By combining cutting-edge perception and control methodologies with robust simulations, the framework demonstrates significant improvements in tracking accuracy, training efficiency, and real-world applicability.

Project Information

  • Category: Computer Vision, Robotics
  • Duration: 12 months
  • Completed: 2025
  • Institution: University at Buffalo
  • Publication: IEEE Robotics & Automation

Technologies Used

Python
PyTorch
AirSim
Unreal Engine
OpenCV
ROS 2

Interested in this project?

If you're interested in learning more about our autonomous UAV tracking system or exploring potential collaborations in computer vision and robotics, please get in touch.

Contact Me