Efficient Violence Detection in Surveillance.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Intelligent video surveillance systems are increasingly deployed in public areas.
• Computer vision and machine learning enable safety monitoring applications from video data.
• Accurate and efficient violent event detection is crucial for public safety.

Purpose of the Study:

• To propose a novel, computationally light architecture for violence detection in video surveillance.
• To enhance the accuracy and efficiency of detecting violent events.

Main Methods:

• A U-Net-like network architecture for spatial feature extraction.
• MobileNet V2 utilized as the encoder for efficient feature extraction.
• Long Short-Term Memory (LSTM) networks for temporal feature extraction and classification.

Main Results:

• The proposed model achieved an average accuracy of 0.82 ± 2%.
• The model demonstrated an average precision of 0.81 ± 3%.
• Effective performance was validated on the complex RWF-2000 real-world security camera footage dataset.

Conclusions:

• The novel architecture offers a computationally efficient solution for violence detection.
• The model achieves competitive accuracy and precision in real-world surveillance scenarios.
• This approach can significantly improve safety monitoring in public spaces.