Context Aware Anomalous Behaviour Detection in Crowded Surveillance
James Leach, PAUL B. SPARKS, Matthew Robertson · 2013
This work addresses the detection of human behavioural anomalies in surveillance. We address in particular the problem of detecting subtle behaviour in a crowded behaviourally heterogeneous surveillance scene. We novel methods of extracting scene context and social context to improve the detection of behavioural anomalies, and in particular permit the detection of subtle behavioural anomalies. Our approach is unsupervised, detecting statistical outliers in a dataset by comparrison of the outlier to the full data. We find that our context aware method performs significantly better than the equivalent method without contextual information. We observe that the use of a contextual information leads to the detection of instances of subtly abnormal behaviour, which otherwise remain indistinguishable from normal behaviour. As a society we have the need to monitor public and pri- vate space in order to prevent criminal behaviour and identify security threats. The scale at which surveillance is undertaken, the density of information in video results in a huge amount of data - the analysis of which using human resources is of- ten prohibitively expensive. The solution is to automate human surveillance (12). Such systems have the potential to monitor a greater amount of data at a lower operating cost and miti- gate the burden on security sta . In order to expose salient behaviour in a video we look for abnormal observations. An abnormal event is one which has a low statistical representa- tion in the training data (9). We motivate our by this defini- tion with emphasis upon contextual information as a method of creating distinct separation between otherwise only subtly dis- tinct behaviours. A good behaviour representation should en- code the dataset in such a way that homogeneous clusters of be- haviour can be segmented from the heterogeneous mass of data. Equally a poor behaviour representation is incapable of measur- ing the distinction between desired subgroups of data. Subtle behaviours provide a greater challenge because the information required to segment them from the greater set is not readily measurable. Subtle behaviours can be handled in the follow- ing two ways; firstly by measuring more relevant information which better segments the data into homogeneous subsets, or secondly by implementing a better suited model which is capa- ble of fitting the nuances of the data domain. In this research we tackle the former point; inspired by work in Scene Mod-