CrimeNet: Neural Structured Learning using Vision Transformer for violence detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models have advanced video violence detection but struggle with high false alarm rates in complex datasets.
• Existing systems often have average precision below 90%, leading to potential system disablement by security personnel.

Purpose of the Study:

• To develop a novel neural network, CrimeNet, that significantly improves violence detection accuracy and minimizes false positives in video surveillance.
• To enhance the robustness and generalizability of violence detection models.

Main Methods:

• A new neural network, CrimeNet, was developed using Vision Transformer (ViT) and Neural Structured Learning (NSL) with adversarial training.
• The model was tested on four challenging binary and multi-class violence-related datasets.
• A generalization study was conducted by training and testing the model across different datasets.

Main Results:

• CrimeNet demonstrated superior performance, outperforming previous methods by a significant margin.
• The model achieved substantial improvements in ROC AUC, ranging from 9.4 to 22.17 percentage points across datasets.
• False positive rates were reduced to near zero, enhancing reliability.

Conclusions:

• CrimeNet offers a significant advancement in video violence detection, drastically reducing false alarms and improving detection accuracy.
• The model exhibits remarkable robustness and generalizability, outperforming competing methods by 12.39 to 25.22 percentage points in generalization tests.