Target Tracking via Incorporating Multi-modal Features

The challenge of visual tracking is to develop a robust target’s appearance model, the core of which involves an appropriate selection and an effective assembly of a cluster of features. In this paper, we propose a novel model to adaptively choose reasonable combination of feature sets to represent the target by employing the multi-kernel ridge regression. This model will update the weights distributions of different kernel groups for feature sets automatically and the regularization parameter’s value of the kernel regression objective function as well. For traditional multi-kernel based algorithm would cost too much time on training model, we develop a very simple and efficient algorithm by adapting feature sets to circulant structure so as to make use of the Fast Fourier Transform (FFT). Thus our algorithm can provide more robust tracking while maintaining real-time effects. To the best of our knowledge, this is the first time the multiple kernel learning algorithms is applied to real-time visual tracking. We evaluate the proposed algorithm on the popular benchmark including 50 image sequences and compare it with 9 state-of-art methods. Implemented in Matlab, the experiment results show that the proposed tracker runs at 45.4 frames per second on an i3 machine and outperforms the state-of-the-art trackers on the benchmark with respect to accuracy. Particularly, the average precision of our algorithm achieves 76.7 % under OPE curve at 20px.