💻 computer science

Suicide Risk Assessment from AI-powered Video Surveillance: An Interpretable Framework for Prevention in Metro Stations

This paper introduces the first interpretable AI framework for Suicide Risk Assessment in metro stations, which integrates person tracking, activity recognition, and spatial-temporal analysis of surveillance video to achieve an 83.2% ROC-AUC on real-world data for early intervention.

Original authors: Safwen Naimi, Wassim Bouachir, Guillaume-Alexandre Bilodeau, Brian Mishara

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Safwen Naimi, Wassim Bouachir, Guillaume-Alexandre Bilodeau, Brian Mishara

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy subway station as a giant, flowing river of people. Usually, everyone moves with a steady, predictable rhythm. But sometimes, a single person might start acting differently—pacing back and forth, staring into the dark tunnel, or lingering dangerously close to the edge. In the past, station staff had to watch hundreds of screens constantly, hoping to spot these subtle signs before a tragedy happened. It's like trying to find a single drop of water that's about to fall off a cliff while watching a whole waterfall.

This paper introduces a new "smart assistant" for subway stations, called SRA-Framework. Think of it not as a mind-reading machine, but as a highly observant, tireless detective that watches the video feeds and asks: "Is this person's behavior, combined with where they are standing, building up a pattern of danger?"

Here is how this system works, broken down into simple steps:

1. The Three Eyes of the System

Instead of just looking for one scary action (like jumping), the system uses three different "eyes" to understand the whole picture:

The Tracker (The Bodyguard): It follows every person like a shadow, keeping track of where they are and how they move, even if the crowd gets thick.
The Actor (The Translator): It recognizes what people are doing. Is the person walking? Standing still? Or looking down the tunnel?
The Map-Maker (The Architect): It understands the station layout. It knows exactly where the "safe zone" is, where the "yellow safety line" is, and where the "danger zone" near the tunnel begins.

2. The "Risk Heatmap" (The Weather Map)

Imagine the subway platform as a map. The system draws a "heat map" over it.

If a person just walks through, the map stays cool (blue).
If a person stands near the edge for a long time, or paces back and forth between the wall and the yellow line, that spot on the map starts to glow red.
The system doesn't just look at one moment; it watches how long the "glow" stays. A quick glance is fine; a long, lingering stare at the danger zone makes the red glow brighter.

3. The "Risk Score" (The Weather Forecast)

The system combines all these clues into a single number, a Risk Score, for each person.

Low Score: The person is just waiting for the train, maybe looking at their phone. No worry.
High Score: The person is doing a "danger dance." They might be:
- Standing on the yellow line.
- Walking back and forth between the wall and the line repeatedly.
- Staring into the tunnel for a long time.
- Spending a lot of time in the "far-end" zone of the platform.

If a person does one of these things, the score might go up a little. But if they do several of them, or do them for a long time, the score shoots up. The system is designed to notice this accumulation of small signs, rather than waiting for a big, obvious crisis.

4. Why "Interpretable" Matters

Many AI systems are "black boxes"—they give an answer but won't tell you why. This system is different. It's like a teacher grading a test and showing you the work:

"This person got a high risk score because they crossed the yellow line 5 times and stared at the tunnel for 30 seconds."
This is crucial because it helps human staff trust the system. They can see the specific reasons for the alarm, rather than just getting a random "danger" signal.

What Did They Find?

The researchers tested this system on real video footage from a Canadian subway station.

The Result: The system was able to distinguish between people who were at risk and those who were just normal commuters with about 83% accuracy.
The Key Insight: The most important clues weren't just "looking at the tunnel." The system found that repetition and location were the biggest factors. Specifically, crossing the yellow line and pacing back and forth were the strongest signals of danger.

The Bottom Line

This paper presents a new tool that helps subway staff spot potential tragedies earlier by watching for patterns of behavior rather than single events. It doesn't guess what someone is thinking; it simply measures how long they stay in dangerous spots and how often they repeat risky movements. By combining computer vision with knowledge from suicide prevention experts, it aims to give human operators a "second pair of eyes" that never gets tired, helping to intervene before it's too late.

Important Note: The paper emphasizes that this system assesses observable risk based on behavior and location. It does not diagnose mental health conditions or claim to know a person's internal thoughts or intent. It is a safety tool designed to flag situations that need human attention.

Technical Summary: Suicide Risk Assessment from AI-powered Video Surveillance

Problem Definition
The paper addresses the challenge of Suicide Risk Assessment (SRA) in metro stations using continuous surveillance video. While existing vision-based systems often rely on simple visual cues (e.g., crossing safety barriers) or continuous human observation, these methods frequently detect incidents too late for effective prevention or suffer from human fatigue. The core problem is to move beyond isolated event detection or direct intent inference. Instead, the authors define SRA as a task requiring the joint reasoning of individual passenger behavior, spatial context, and temporal dynamics. The goal is to assess a time-varying risk score for each passenger based on the accumulation of observable behavioral and spatial cues, rather than making a clinical diagnosis of suicidal intent.

Methodology: The SRA-Framework
The authors propose SRA-Framework, the first interpretable pipeline designed to assess suicide risk by aggregating evidence over time. The framework operates through a unified pipeline consisting of five sequential modules:

Human Detection, Tracking, and Pose Estimation:
- Uses a pre-trained YOLOx detector to identify "person" classes.
- Employs ByteTrack for multi-object tracking in image space.
- Applies HRNet to extract 2D body keypoints within bounding boxes.
- Projects tracked positions into a platform-referenced coordinate system using a homography matrix, enabling consistent spatial analysis across time.
Human Activity Recognition:
- Utilizes SSTAR, a skeleton-based spatio-temporal action recognition model pre-trained on the ARMM dataset.
- Specifically detects three passenger actions: 'LookTunnel', 'Walk', and 'Stand'.
- These labels are weighted by their spatial context and duration.
Platform Semantics and Zones Construction:
- Uses YOLOv8n for semantic segmentation to partition the platform into relevant spatial zones based on expert consultation:
  - Zone A (Wall-Proximal): Near the platform wall.
  - Zone B (Yellow-Line Proximal): The transitional safety zone.
  - Zone C (Far-End Zone): The area immediately adjacent to the tunnel/edge.
- This geometric modeling allows for the detection of specific patterns like back-and-forth movement between the wall and the yellow line.
Position Risk Heatmap Generation:
- Generates individual heatmaps via density accumulation and Gaussian smoothing of 2D trajectory points.
- Aggregates these into a global Position Risk Heatmap ( $H(x,y)$ ) by averaging trajectories of known at-risk individuals. This captures shared spatial patterns and persistent environmental factors.
Suicide Risk Assessment (Inference):
- Extracts eight interpretable risk indicators for each individual $p$ $p$ over a temporal window $W_t$ $W_{t}$ :
  1. Position Risk Score ( $PR_p$ )
  2. Walk/Stand on the yellow line ( $Cr_p$ )
  3. Number of yellow line crossings ( $NCr_p$ )
  4. Time spent on the yellow line ( $TY_p$ )
  5. Longest continuous segment on the yellow line ( $LY_p$ )
  6. Back-and-forth count between wall and yellow line ( $BF_p$ )
  7. Look Tunnel action ( $LT_p$ )
  8. Entry into the far-end zone ( $E_p$ )
- These indicators form a feature vector fed into an XGBoost supervised model to output a continuous risk score $R_p \in [0, 1]$ .

Key Contributions

Formalization of SRA: The work defines SRA as a distinct computational task focused on assessing risk levels based on accumulated behavioral and spatial evidence, rather than predicting intent or labeling individuals.
Interpretable Pipeline: The authors introduce the first end-to-end framework that integrates tracking, activity recognition, semantic segmentation, and risk modeling to produce continuous, individual-level assessments without relying on appearance cues or personal attributes.
Alignment with Interdisciplinary Findings: The risk indicators are explicitly derived from prior psychological and behavioral studies (e.g., Mishara et al.), ensuring the system reflects established patterns such as repeated pacing and prolonged presence near the platform edge.
Real-World Evaluation: The framework is evaluated on ethically approved, real-world surveillance data from a Canadian metro system, providing both quantitative performance metrics and qualitative interpretability analyses.

Experimental Results
The framework was evaluated on a dataset comprising 66 "at-risk" video recordings (pre-attempt) and 56 "control" recordings, yielding 256 individual instances (66 at-risk, 190 control).

Performance: The fully automatic configuration achieved an ROC-AUC of 83.2% ( $\pm 0.095$ ).
Sensitivity/Specificity: At a fixed threshold of 0.2 (chosen to favor early detection), the system achieved a sensitivity of 63.1% and a specificity of 90.2%.
Upper Bound Analysis: When detection/tracking and activity recognition modules were replaced with ground-truth annotations, the ROC-AUC improved to 91.9% and 89.3% respectively, indicating that errors in early perception modules contribute significantly to the performance gap.
Interpretability: Feature importance analysis (via XGBoost and SHAP values) confirmed that indicators related to direct interaction with the yellow line (crossing, time spent, back-and-forth movement) were the most dominant predictors, aligning with prior behavioral research.

Significance and Claims
The paper claims that its primary significance lies in bridging the gap between AI-based surveillance and interdisciplinary suicide prevention research. By formalizing SRA as a context-aware, long-term risk assessment problem, the framework moves beyond simple anomaly detection or intrusion alerts. The authors emphasize that the system is designed to be interpretable, allowing human operators to understand why a risk score is high (e.g., due to specific spatial patterns and temporal persistence) rather than relying on a "black box" prediction. This approach aims to support human operators in real-time monitoring while adhering to ethical constraints, avoiding the inference of psychological states or the use of personal attributes. The work highlights the complexity of suicide risk assessment and opens new directions for developing interpretable AI systems for social good.