Geometric Layout Based Graphical Model for Multi-Part Object Tracking

Overview & Synopsis

This work puts forth a probabilistic graphical framework to track unoccluded objects undergoing large out of image plane rotations and/or presenting large scale variations in video sequences. The proposed scheme incorporates measurements from an ensemble of local patch trackers and inter-patch geometric layout to arrive at a sample based approximation of the state posterior. Following this, the geometric layout is updated online using the Iterative Conditional Estimation technique. These steps are iterated until convergence to arrive at the final state posterior. In contrast to offline training based schemes the proposed framework imposes no prior on the geometric layout and instead relies on online update of the geometric layout, thus broadening the scope of usage. Amongst other advantages, the scheme implicitly estimates the scale of the target and also adapts to varying target appearances to enable tracking under a fair degree of out of the image plane rotations. The tracking abilities of this scheme is put to test on several challenging videos with scale changes, out of the image plane rotations, illumination changes and motion jerks. Whereever possible qualitative comparisons are facilitated using videos from standard databases.