Towards Building a Visual Behaviour Analysis Pipeline for Suicide Detection and Prevention

Area of Science:

• Computer Vision
• Artificial Intelligence
• Behavioral Science

Background:

• Deep learning advances human behavior analysis via video, crucial for intelligent surveillance.
• Current systems lack effective suicide prevention capabilities, especially for early crisis detection.
• Privacy concerns limit datasets for detecting pre-suicidal behaviors.

Purpose of the Study:

• Develop an intelligent visual processing pipeline for real-time human behavior analysis.
• Address the challenge of detecting crisis behaviors before suicide attempts.
• Facilitate early intervention and prevention at suicide hotspots.

Main Methods:

• Proposed a modular pipeline including pedestrian detection, tracking, and pose estimation.
• Developed a novel 2D skeleton-based action recognition algorithm using a two-branch network and stacked LSTM.
• Integrated a logical layer informed by human coding studies to recognize pre-suicide indicators.

Main Results:

• Achieved strong performance in action recognition on public and private datasets.
• Demonstrated effectiveness in recognizing complex behavior patterns from limited data.
• Validated the pipeline's capability in capturing crisis behavior indicators using simulated surveillance footage.

Conclusions:

• The proposed system effectively detects pre-suicide behaviors for early intervention.
• This AI-driven approach offers a promising solution for suicide prevention in visual surveillance.
• The modular design and novel algorithms advance the application of AI in public safety.