Articulated human tracking and behavioural analysis in video sequences

Abstract

Recently, there has been a dramatic growth of interest in the observation and tracking of human subjects through video sequences. Arguably, the principal impetus has come from the perceived demand for technological surveillance, however applications in entertainment, intelligent domiciles and medicine are also increasing. This thesis examines human articulated tracking and the classi cation of human movement, rst separately and then as a sequential process. First, this thesis considers the development and training of a 3D model of human body structure and dynamics. To process video sequences, an observation model is also designed with a multi-component likelihood based on edge, silhouette and colour. This is de ned on the articulated limbs, and visible from a single or multiple cameras, each of which may be calibrated from that sequence. Second, for behavioural analysis, we develop a methodology in which actions and activities are described by semantic labels generated from a Movement Cluster Model (MCM). Third, a Hierarchical Partitioned Particle Filter (HPPF) was developed for human tracking that allows multi-level parameter search consistent with the body structure. This tracker relies on the articulated motion prediction provided by the MCM at pose or limb level. Fourth, tracking and movement analysis are integrated to generate a probabilistic activity description with action labels. The implemented algorithms for tracking and behavioural analysis are tested extensively and independently against ground truth on human tracking and surveillance datasets. Dynamic models are shown to predict and generate synthetic motion, while MCM recovers both periodic and non-periodic activities, de ned either on the whole body or at the limb level. Tracking results are comparable with the state of the art, however the integrated behaviour analysis adds to the value of the approach.