A lightweight convolutional neural network architecture for violence detection in video sequences.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Escalating violent incidents in public venues necessitate efficient real-time detection systems.
• Processing high-dimensional video data for violence detection is computationally intensive and complex.
• Existing methods struggle with spatiotemporal variations and illumination inconsistencies.

Purpose of the Study:

• To develop a computationally efficient and accurate real-time violence detection framework.
• To significantly reduce computational overhead without compromising classification accuracy.
• To enable deployment on resource-constrained hardware for real-world surveillance.

Main Methods:

• Developed a lightweight deep convolutional neural network (CNN) architecture based on MobileNetV2.
• Optimized the CNN using depthwise separable convolutions and inverted residual bottlenecks.
• Preprocessed video frames (224x224 resolution, normalization, augmentation) for enhanced generalization.
• Trained and evaluated the model on the Real-Life Violence Situations Dataset (RLVSD) and Hockey Fight Dataset (HFD).

Main Results:

• Achieved 97% accuracy on RLVSD and 94% accuracy on HFD.
• Demonstrated superior precision, recall, and F1-score compared to conventional CNN architectures.
• Confirmed substantial efficiency improvements, enabling real-time inference on resource-constrained hardware.

Conclusions:

• Optimized lightweight CNN architectures can achieve high-accuracy violence detection with reduced computational cost.
• The proposed framework is highly deployable for real-world surveillance systems.
• Future work will explore temporal feature integration and cross-domain adaptability.