‘Surprise’ Detection in Human Pattern-of-Life Behaviour

In smart city data analytics, the observation of human behavioural characteristics from automated sensor analysis is a useful tool to help identify common sub-clusters and anomalous activity relative to the population’s pattern-of-life dynamic. Avoiding the use of video and mobile phone active applications, because of ethical and privacy concerns, we are more interested in what information can be extracted from lower sensing modalities, such as basic tracking data as obtained from time course GPS signals for example. However, trajectory data has its own technical problems; primarily sparsity, non-uniform sampling, usually unlabelled and sporadic sensor data location errors. In this paper we analyse basic track data obtained from consenting subjects using an approach which combines both a direct time series modelling framework, and a probabilistic framework; the latter derived from an acknowledgement that the residual errors in modelling are never Gaussian distributed for real world data due to our uncertainties in data, models and observations. We construct a hybrid dissimilarity representation of the problem to visualise anomalous individuals without needing extreme value statistics (i.e.. behaving differently to the common consensus trajectories) which also reveals preliminary insights into population sub-clusters. In disaster management situations, or pre-disaster planning, this information can help in quickly identifying dangerous behaviour and predicting dangerous behavioural sub-grouping in population movement prompting fast and effective action.