Contextual information based anomaly detection for multi-scene aerial videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Surveillance Technology

Background:

• Aerial video surveillance using Unmanned Aerial Vehicles (UAVs) is crucial for diverse applications like wildlife monitoring and disaster management.
• Manual analysis of surveillance videos is time-consuming and subjective, necessitating automated solutions.
• Existing methods often fail with dynamic backgrounds in UAV footage due to a lack of suitable datasets.

Purpose of the Study:

• To address the limitations in analyzing dynamic aerial surveillance videos.
• To introduce a novel multi-scene dataset for Unmanned Aerial Vehicle (UAV)-based anomaly detection.
• To develop a Computer Aided Decision (CAD) support system for accurate anomaly detection in UAV videos.

Main Methods:

• A new dataset with frame-level annotations for multi-scene aerial anomaly detection was created.
• A novel CAD system was developed, integrating contextual, temporal, and appearance features.
• A unique feature descriptor was designed to capture multi-scene contextual information.

Main Results:

• The proposed system demonstrated competitive performance against state-of-the-art methods.
• An Area Under the Curve (AUC) of 0.712 was achieved on the new UAV anomaly detection dataset.
• The system effectively utilizes contextual, temporal, and appearance features for anomaly detection.

Conclusions:

• The developed dataset and CAD system significantly advance UAV-based anomaly detection.
• The system's ability to analyze dynamic, multi-scene videos offers a robust solution for surveillance.
• This work provides a foundation for future research in intelligent aerial surveillance systems.