Summary: Autonomous UAVs are increasingly deployed in diverse environments such as urban areas, forests, construction sites, and agricultural fields. However, object detection models are typically trained offline using fixed datasets and often experience significant performance degradation when operating in unseen environments. Retraining the model from scratch whenever new data becomes available is computationally expensive and impractical for real-world deployment. This project investigates the use of Continual Learning (CL) to enable UAV object detection models to incrementally learn from new environments while preserving previously acquired knowledge. The project will compare several continual learning techniques and evaluate their effectiveness in mitigating catastrophic forgetting under realistic UAV deployment scenarios.
Objectives
- Develop a UAV object detection system using a modern detector (e.g., YOLO or RT-DETR).
- Implement and compare several continual learning algorithms such as Experience Replay, Elastic Weight Consolidation (EWC), and Learning without Forgetting (LwF).
- Evaluate the impact of continual learning on object detection accuracy, catastrophic forgetting, computational cost, and memory requirements.
- Investigate the trade-offs between model adaptability and knowledge retention across multiple UAV datasets.
- Provide recommendations for selecting continual learning strategies suitable for real-world UAV applications.
Expected Outcomes. The project will produce:
- A complete continual learning framework for UAV object detection.
- Comparative evaluation of multiple continual learning algorithms.
- Quantitative analysis of catastrophic forgetting.
- Performance comparison across different environments and datasets.
- Design recommendations for deploying continual learning in autonomous UAV systems.
- A comprehensive dissertation discussing methodology, experiments, limitations, and future work.
Key Requirements
Students should have:
- Good Python programming skills.
- Knowledge of deep learning (PyTorch preferred).
- Basic understanding of computer vision.
- Familiarity with CNN-based object detection.
- Willingness to work with GPU computing.
- Interest in autonomous systems or robotics.
Recommended knowledge: YOLO, PyTorch, OpenCV, basic ML