End-to-End Deep One-Class Learning for Anomaly Detection in UAV Video Stream.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicle (UAV) surveillance is increasing globally.
• Anomaly detection in surveillance often lacks sufficient data for normal events, necessitating unsupervised learning.
• Existing methods struggle with unsupervised anomaly detection using only normal event data.

Purpose of the Study:

• To develop a novel end-to-end architecture for unsupervised anomaly detection in drone surveillance.
• To generate optical flow images and extract compact spatio-temporal features from UAV imagery.
• To address the challenge of limited normal event data in unsupervised learning scenarios.

Main Methods:

• Proposed an end-to-end generative learning architecture for unsupervised anomaly detection.
• Implemented a custom loss function comprising reconstruction loss (Rl), generation loss (Gl), and compactness loss (Cl).
• Utilized background subtraction on optical flow images to mitigate UAV motion effects.

Main Results:

• The proposed method successfully generates optical flow and extracts spatio-temporal features.
• Achieved an Area Under the Curve (AUC) of 85.3% on the mini-drone video dataset.
• Demonstrated superior performance compared to existing anomaly detection techniques.

Conclusions:

• The developed generative learning method is effective for unsupervised anomaly detection in drone surveillance.
• The custom loss function and motion compensation technique contribute to improved performance.
• The approach offers a promising solution for real-world surveillance applications with limited anomaly data.