PCA-Based Estimation of Velocity Feature for Visual Tracking

The performance of visual tracking mainly depends on observation models and search methods. To stabilize the tracker, an algorithm for estimating velocity feature based on principal component analysis (PCA) is proposed. The proposed algorithm calculates the velocity feature using PCA by the states of the object in the previous k frames. In addition, integrating color and motion cues for visual tracking is implemented with a layered particle filter which consists of two stages of importance sampling. At the first stage, the state of object is crudely estimated with color kernel histogram in RGB color space. The velocity features are used to precisely calculate the state of object in the second stage. The experimental results show that the proposed algorithm has stronger ability to resist to interference of noise, and significantly improves the tracking accuracy in comparison with tracker based on the standard particle filter with a single color cue.