A high frequency of violent harmful behaviour occurs within forensic mental health services. The highest level of secure hospital cares for patients who are most at risk of harm to others and often also themselves. In the UK, monitoring and management of risk of harm typically involves close proximity observation and monitoring by designated staff, which can result in patients experiencing invasion of privacy and holds potential to escalate harmful behaviour. For aggression monitoring, computer vision based surveillance approaches have been used in the US Prison system (e.g. Ellis, 1999) and between groups of actors in a study designed to simulate a prison yard (Chang, Krahnstoever, Lim & Yu, 2010) but the real-world application of intelligent computer vision systems to alerting aggressive behaviour that overcome issues of data privacy and identity is yet to be explored.The aim of this study is to develop the first anonymised approach to measurement of aggressive behaviour based around 3D camera technology, with a view to developing an action recognition approach to generate alerts for aggressive/harmful behaviour. The observation system was installed in a ward of a high secure hospital in the UK and provided pseudo-continuous 3D observation of a known closed-set of at-risk patients. Individual antecedent behaviours to an aggressive incident were analysed for developing automatic early alerts to aggression. This approach offers scope as a resource for building long-term observational models of behaviour that could inform future care planning as well as monitoring clinical effectiveness, so enhancing an understanding of violent and risk-related behaviours.