On the Behaviour of Body Tracking with the Annealed Particle Filter in Realistic Conditions

Articulated full-body tracking of high-dimensional 3D human figures is currently an active area of research. However, the vast majority of work has focused on tracking a person in controlled environments similar to dedicated marker-based motion-capture facilities. This means that many real-world considerations are ignored, such as background clutter, tracking various people of different builds, poor contrast and low-level algorithm failures. These types of real-world conditions are typical for many vision applications that could benefit significantly from the detailed posture description that full-body tracking provides, such as action recognition or multi-person tracking. Hence this paper evaluates the behaviour of the annealed particle filter (APF – a widely-cited body tracking algorithm) with the goal of identifying and resolving the cause of failures in tracking under realistic conditions. A loose body model is employed to specifically allow for the possibility to track people of different builds. Findings show that much of reason for tracking failures can be traced to the observationevaluating function in conjunction with the loose body model, and an improved observation likelihood function is proposed that substantially reduces the occurrence and severity of failures.