Safety After Dark: A Privacy Compliant and Real-Time Edge Computing Intelligent Video Analytics for Safer Public...

Area of Science:

• Computer Science
• Artificial Intelligence
• Transportation Security

Background:

• Public transportation faces increasing security threats, particularly violence.
• Real-time detection of violence is critical for passenger safety and system operations.

Purpose of the Study:

• To develop and evaluate an advanced AI solution for identifying unsafe behaviors in public transport.
• To enhance the safety and passenger experience in urban transit systems.

Main Methods:

• Utilized deep learning action recognition models, specifically a Temporal Pyramid Network (TPN).
• Integrated NVIDIA DeepStream SDK, AWS DirectConnect, edge computing, ONNXRuntime, and MQTT for an accelerated pipeline.
• Trained the TPN model on a curated dataset combining open-source and simulated live trial videos.

Main Results:

• Achieved 97% validation accuracy after incorporating live simulated data into training.
• Demonstrated impressive real-world performance during an 8-week trial at Wollongong Train Station.
• Recorded a 23% false-positive rate and correctly identified 30 true-positive violence incidents, with 6 false negatives, particularly in adverse weather.

Conclusions:

• The AI system shows significant potential for enhancing public transport safety through accurate violence detection.
• Continuous retraining capabilities ensure adaptability to diverse real-world conditions.
• Further data, especially for adverse weather, is needed to improve model robustness.