Efficient Human Violence Recognition for Surveillance in Real Time.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human violence recognition is crucial for public safety and surveillance.
• Existing methods often prioritize precision over real-time efficiency.
• There is a need for practical, efficient violence detection models.

Purpose of the Study:

• To develop an effective and efficient model for real-time human violence recognition.
• To address the limitations of existing precision-focused approaches.
• To create a robust system applicable to real-world surveillance scenarios.

Main Methods:

• A novel three-module model: Spatial Motion Extractor (SME), Short Temporal Extractor (STE), and Global Temporal Extractor (GTE).
• SME extracts regions of interest, STE captures rapid movement dynamics, and GTE identifies long-term temporal features.
• The model was evaluated for efficiency, effectiveness, and real-time performance.

Main Results:

• The proposed model demonstrated high efficiency on Hockey, Movies, and RWF-2000 datasets.
• Superior effectiveness was achieved on the newly created VioPeru dataset, specifically designed for real-world surveillance.
• The model's real-time applicability was validated through rigorous testing.

Conclusions:

• The developed model offers a significant advancement in efficient and effective real-time human violence recognition.
• It provides a practical solution for video surveillance systems, enhancing public safety.
• The VioPeru dataset serves as a valuable benchmark for future research in this domain.