Abstract:
The importance of behaviour analysis and activity recognition systems continue to increase with the increasing demand and deployment of video surveillance systems. Motion trajectories provide rich spatio-temporal information about an object’s activity. In this article, we present a supervised feature extraction and multivariate modelling approach for motion-based behaviour recognition and anomaly detection. In the proposed motion learning system, trajectories are treated as time series and modelled using modified DFT-based coefficient feature space representation. We employ supervised dimensionality reduction using Local Fisher Discriminant Analysis to enhance the feature space representation of trajectories. A modelling approach, referred to as multivariate m-mediods, is proposed that can cater for the presence of multivariate distribution of samples within a given motion pattern. A hierarchical indexing of mediods and retrieval approach is presented to improve the efficiency of proposed classifier. Our proposed techniques are validated using variety of simulated and complex real-life trajectory datasets.
