Joint target tracking and recognition using view and identity manifolds

Vijay Venkataraman, Guoliang Fan, Liangjiang Yu, Xin Zhang, Weiguang Liu, Joseph P. Havlicek

Abstract: We propose a new concept of identity manifold for automated target tracking and recognition (ATR) that captures both inter-class (e.g., between tanks and armored cars) and intra-class (e.g., between different tanks) variability of target appearances (e.g., shapes). A hemisphere-shaped view manifold is also involved for mutli-view target modeling. Combining the two continuous-valued manifolds via nonlinear tensor decomposition gives rise to a new generative model that can be learned from a small training set. This model can not only deal with arbitrary view/pose variations by tracking along the view manifold, but also interpolate the appearance of an unknown target along the identity manifold. The proposed model is examined based on the recently released SENSIAC ATR database, and the experimental results confirm the usefulness of this generative model.