Visual tracking is defined as a process in which a video camera is used to recover the location and motion of a target, over a given time period, as it moves from its starting point to its end point. Visual tracking is an important scientific problem; the human visual system is capable of tracking moving objects in a wide variety of situations. Visual tracking is also the key to a number of important practical applications such as video surveillance, 3D shape recovery and target restoration.
Many powerful visual tracking techniques exist. Most of them are built using model estimation concepts, such as Kalman filters or Sequential Monte Carlo methods, particularly particle filters. A significant advantage that particle filters have over Kalman filters is that particle filters can represent ambiguity in the target’s position, whereas Kalman filters can only represent a single hypothesis at a time. The multiple hypotheses maintained by particle filters can be viewed as physical particles sprinkled over the image plane, and are often shown as spots of different colours or intensity.
While many effective tracking algorithms have been based on the particle filter, problems remain. Occlusion arises when another object, with different features, e.g. colour, comes between the camera and target. Camouflage occurs when an object with similar features lies behind the target and makes the target invisible from the camera’s point of view. Either of these events can cause a tracker become dissociated from its target, so that the data it produces is unrelated to the target’s behaviour.
Previous approaches to these problems have sought to keep the tracker’s particle set more tightly focused on the target. This is done by incorporating more information about the target and/or the environment it is moving through. Multiple models of motion might be used to predict the target’s future location, more detailed texture and colour cues might be used to better model its appearance, or areas of the environment in which the target is more likely to appear might be identified. These approaches, however, do not entirely remove the problems of occlusion and camouflage. As a result, we would argue that the solution lies not in avoiding, but in detecting and reacting to these disruptive events.
In the present paper we focus on the detection of occlusion and camouflage during
particle filter-based tracking. We use a clustering algorithm, the Expectation Maximization
algorithm, to investigate the distribution of the particles representing a given target
during the events of occlusion and camouflage. This clustering process provides
vital information about fluctuation in the behaviour of particle sets, which is assessed
using a process-behaviour chart. The process-behaviour chart will alert the tracker
of the occurrence of occlusion or camouflage. The information produced by the
process-behaviour chart is then used by the tracker to map out the boundary of
the interfering object, providing valuable information about the viewed environment.
