Object tracking by adaptive feature extraction

Tracking objects in the high-dimensional feature space is not only computationally expensive but also functionally inefficient. Selecting a low-dimensional discriminative feature set is a critical step to improve tracker performance. A good feature set for tracking can differ from frame to frame due to the changes in the background against the tracked object, and due to an on-line algorithm that adaptively determines a advantageous distinctive feature set. In this paper, multiple heterogeneous features are assembled, and likelihood images are constructed for various subspaces of the combined feature space. Then, the most discriminative feature is extracted by principal component analysis (PCA) based on those likelihood images. This idea is applied to the mean-shift tracking algorithm [D. Comaniciu et al., June 2000], and we demonstrate its effectiveness through various experiments.