On the Construction of Proposal Distributions based on Linear Estimation for Visual Tracking

Abstract

Linear estimation based sequential importance sampling methods for particle filters are proposed that can be used to track an object in a video sequence, especially to track an object with rapid motion changes which cannot be well modeled by a prior model. First a linear least-squares estimation (LSE) is used to construct an importance function based on observations, and then it is extended to a robust linear estimation (RLE). The LSE does not require solving the normal equations at every time with a simple assumption; it reduces to matrix multiplication. In addition the RLE only requires solving linear systems with 4x4 matrices several times; numerical search is not required. An experiment with a real video sequence involving a rapidly moving object is made. The results show that the proposed sampling methods outperform the prior model based sampling method when the rapid motion change of the object happens. In addition it is shown the RSE based sampling method successfully track the object even if the LSE based sampling methods fails because of outliers.