Hierarchical active shape model with motion prediction for real-time tracking of non-rigid objects

Abstract

Video tracking systems generally deal with non-rigid objects with various shapes and sizes. This often results in a poor match of an initial model with the actual input shape, and consequently causes the failure of tracking. The robustness of the active shape model (ASM) enables video tracking systems to deal with such unpredictable inputs. The iterative nature of the ASM, however, makes real-time implementation difficult. A novel ASM-based real-time tracking method with particular relevance to non-rigid objects is proposed. The proposed tracking system adopts a hierarchical approach to reduce computational loads and increase immunity to noise. In order to make the system operate in real-time, a novel prediction approach is proposed that significantly reduces the number of iterations. In the sequential images, the initial feature points have been estimated using a block-matching algorithm. Kalman filtering has also been applied for increasing accuracy of the motion estimation. The proposed hierarchical, prediction-based approach is proven to outperform the existing methods in the sense of both tracking performance and convergence speed.