Real-time violence detection and localization through subgroup analysis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Surveillance Systems

Background:

• Existing violence detection (VD) methods struggle with real-world surveillance data, lacking localization and social context.
• Timely human intervention is critical, but current VD systems are often inadequate for complex scenarios.

Purpose of the Study:

• To develop an adaptable add-on module for VD that integrates subgroup recognition and tracking.
• To enhance real-time VD systems with improved social awareness and localization of violent incidents.

Main Methods:

• Proposed a novel approach to integrate subgroup tracking into existing VD models.
• Developed a system to recognize and track multiple subgroups across video frames.
• Integrated localization capabilities to identify groups involved in violent events.

Main Results:

• The method improves social awareness in real-time VD by localizing individuals involved in violence.
• Achieved 91.3% accuracy on the SCFD dataset and 87.2% on the RWF-2000 dataset.
• Demonstrated practical utility and generalization to unseen datasets, performing close to state-of-the-art.

Conclusions:

• The proposed subgroup integration enhances VD systems, providing crucial localization and social dimension information.
• This adaptable module offers a promising advance for early violence detection in surveillance.
• The method's efficiency and generalization capabilities highlight its practical applicability in real-world scenarios.