Object tracking using distribution fields with correlation coefficients

Abstract

A real-time object tracking method based on distribution field (DF) constructs with correlation coefficients is proposed to solve the drawbacks of local search and poor real-time performance exhibited by traditional DF tracking methods. With the goal of adapting to complex environments and changes in tracking speed, we propose an algorithm based on DFs and global searching by dense sampling. First, we use the DFs to construct an appearance model that functions as a target descriptor in the particle filter framework, allowing dynamic updating of the appearance model. Then, we measure the similarity using correlation coefficients based on fast Fourier transforms (FFTs) instead of the L1-norm of DFs to reduce the time complexity, overcome the drawback of randomness when using sparse sampling, and avoid falling into local optima from the gradient descent used in traditional DF methods. The results of experiments show that our proposed algorithm not only performs in real time but is also more robust for a variety of complex environments than those of six state-of-the-art algorithms on eight challenging video sequences.