Context-Aware Motion Detection in Dynamic UAV Video: A Comparative Evaluation of Deep Learning and Unsupervised Models
Applied Technology- Grade:
- 11
- Teacher:
- Kevin Nieves Pichardo
With the rising relevance of autonomous tracking in real-time, Unmanned Aerial Vehicles (UAVs) play a critical role in surveillance and monitoring. Current UAV motion detection approaches face many limitations. Traditional models assume static cameras while deep learning models require large labeled datasets with high computational costs, posing concerns over effectiveness and adaptability. An unsupervised context-aware tensor-based model addresses these limitations by modeling spatial and temporal relationships without labeled data through penalty-based regularization to improve consistency and reduce noise. We compared this proposed model with state-of-the-art unsupervised models like Robust Principal Component Analysis (RPCA) and Probabilistic Robust Matrix Factorization (PRMF) and a pre-trained deep learning model DeepLab V3+ on identical UAV video sequences for evaluation fairness. We conducted an ablation study on the proposed model to analyze influence of weighted penalty parameters and acquire the best-performing configuration. The optimized proposed model achieved a precision of 0.90, recall of 0.93, and F1 of 0.91. DeepLab V3+ had higher precision (0.95), but much lower recall (0.68), reducing F1 to 0.79. RPCA and PRMF achieved F1s of 0.74 and 0.78 respectively. The superior F1 of the proposed model suggests that an unsupervised context-aware tensor-based motion detection outperforms classical and deep learning approaches while maintaining computational efficiency and without reliance on labeled data, proving its ability to serve as a robust and scalable solution for real-time UAV motion detection applications like search-and-rescue, surveillance, and environmental monitoring.
The problem statement is excellent and demonstrates that the researchers have a real understanding of the research domain. The requirements for the proposed solution are clearly defined: the model must handle dynamic camera motion, reduce background noise false positives, improve detection coverage (recall), and do so without requiring labeled training dat a.