It is estimated that approximately 30% of all suicides occur in public places, and means restriction is an effective approach to preventing suicide. CCTV cameras have been used in a range of settings to identify when a suicide attempt is in progress, although tragically this is often too late to intervene. Identifying behaviours preceding a suicide attempt, and using artificial intelligence (AI) for the automatic detection of these behaviours, has been proposed as a promising avenue for suicide prevention research.
Existing knowledge from international literature about behaviours preceding a suicide in railway settings was synthesised with our own analysis of CCTV data. Behaviours which were unique to individual locations, as well as those that were common or analogous across settings, were observed. A framework for decomposing these behaviours into features such as semantic/contextual location, physical location, and trajectory was developed, allowing complex behaviours to be described in a framework suitable for automatic detection.
Building on previous research at the coastal park setting, we evaluated a computer vision analysis pipeline to generate preliminary evidence for its effectiveness at detecting behaviours across different contexts. Individual modules were validated to have high accuracy across settings, without apparent bias related to demographics. Preliminary analysis indicates a subset of crisis behaviours could be reliably identified with this approach.
Beyond the technical aspects of this research, we also explored the acceptability of this type of suicide.