Evaluating a Human Detection Model in a Behaviour Analysis Pipeline for Suicide Prevention

Area of Science:

• Computer Vision
• Artificial Intelligence
• Public Health

Background:

• Suicides in public areas, particularly railways, cause significant distress to witnesses and staff.
• Pre-suicidal behaviors may be detectable using automated systems.
• Early detection offers a critical window for intervention.

Purpose of the Study:

• To assess the performance of a human detection model for identifying individuals exhibiting pre-suicidal behaviors at railway stations.
• To analyze the model's accuracy in pedestrian detection as a foundational step for a broader suicide prevention project.

Main Methods:

• Analysis of closed-circuit television (CCTV) footage from two railway stations over a 24-hour period.
• Manual annotation of footage to determine true positives, false positives, and false negatives.
• Computation of performance metrics including sensitivity, precision, and F1 score.

Main Results:

• The human detection model demonstrated variable performance across stations.
• Station A: Sensitivity 0.73, F1 score 0.84.
• Station B: Sensitivity 0.48, F1 score 0.65.
• Identified root causes for false negatives include variations in body posture and visual occlusion.

Conclusions:

• The pedestrian detection model is adequate but its effectiveness is camera-view dependent.
• Findings provide insights for improving the model's accuracy in detecting pre-suicidal behaviors.
• Enhanced detection capabilities can support timely interventions, potentially reducing suicide attempts and bystander trauma.