A Competitive Attentional Approach to Mitigating Model Drift in Adaptive Visual Tracking
Sebastien Wong, Adam Gatt, David A. Kearney, Anthony Milton, V. Stamatescu · 2014
A critical issue for adaptive visual tracking is that of model drift, which occurs when the state space of the object of interest is polluted by observations that should have been attributed to background clutter. One approach to mitigating model drift in adaptive feature-learning visual tracking systems is to introduce prior information about the object of interest, such as key-frames. We propose an alternative solution to mitigating model drift, which is to track everything including sources of clutter and then assign observations to the tracks that best describe the observations. We demonstrate that by having multiple single target trackers (Shape Estimating Filters) that interact in a competitive attentional framework, observations from clutter (objects that are not of interest) can be explained-away allowing each tracker to focus its attention on its object of interest.