Bus Violence: An Open Benchmark for Video Violence Detection on Public Transport

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Automatic violence detection in public spaces using AI is challenging due to generalization issues.
• Current AI algorithms require extensive annotated data and perform poorly on unseen scenarios.

Purpose of the Study:

• Introduce and release the "Bus Violence" benchmark, a large-scale video dataset for violence detection on public transport.
• Analyze the performance of state-of-the-art violence detectors on this new benchmark.

Main Methods:

• Collected a large-scale dataset of video clips simulating violent actions on a moving bus with varying conditions.
• Evaluated pre-trained video violence detectors on the "Bus Violence" benchmark.

Main Results:

• State-of-the-art detectors showed moderate performance on the "Bus Violence" dataset.
• Performance drop indicates generalization difficulties of current AI models in this specific use case.

Conclusions:

• The "Bus Violence" benchmark is crucial for developing specialized AI models for public transport violence detection.
• Further research is needed to improve the generalization capabilities of AI-based violence detection systems.