In this paper, a visual object tracking algorithm based on the Kalman particle filter (KPF) is presented. The KPF uses the
Kalman filter to generate sophisticated proposal distributions which greatly improving the tracking performance.
However, this improvement is at the cost of much extra computation. To accelerate the algorithm, we mend the
conventional KPF by adaptively adjusting the number of particles during the resampling step. Moreover, in order to
improve the robustness of tracker without increasing the computational load, another two modifications is made: firstly,
the covariance matrix of Gaussian noise in the dynamic model is dynamically updated according to the accuracy degree
of the prediction. Secondly, the similarity measurement is performed by a scheme that adaptively switches the likelihood
models. Experimental results demonstrate the efficiency and accuracy of the proposed algorithm.