Towards a data-driven adaptive anomaly detection system for human activity

Research in the field of ambient intelligence allows for the utilisation of different computational models for human activity recognition and abnormality detection to promote independent living and to improve the quality of life for the increasing ageing population. The existing monitoring systems are not adaptive to the overly changing human behavioural routine leading to a high rate of false predictions. An adaptive system pipeline is proposed in this paper for adapting to changes in human behaviour based on data ageing and data dissimilarity forgetting factors. The forgetting factor feature allows adaptation of the model to the current routines of an individual while forgetting outdated behavioural patterns. The data ageing forgetting factor discard old behavioural routine based on the age of the activity data while in the data dissimilarity approach, this is achieved by measuring the similarity of the activity data. Behaviour modelling is achieved using an ensemble of novelty detection models termed as Consensus Novelty Detection Ensemble consisting of One-Class Support Vector Machine, Local Outlier Factor, Robust Covariance Estimation and Isolation Forest. The proposed approach is data-driven and environment-invariant, making it feasible for deployment in heterogeneous environments. A comparative analysis carried out with other abnormality detection models for human activities across two datasets shows that the proposed approach achieved better results.