To solve the problem of the exponential growth of the Gaussian mixture components for estimating the state of the non-Gaussian system with the Gaussian mixture model, a cluster-fusion algorithm based on the similarity distribution criterion was proposed. According to that criterion, the Gaussian components were then clustered into different Gauss clusters based on the optimal confidence interval, derived by minimizing the extended integral square error cost function. Meanwhile, to avoid the reuse of the cross components, the local nearest neighbor approach was introduced to re-allocate these cross ones. Then, the components in the clusters were merged by the multielement mergence method to keep with the unbiased property, which can decrease the number of the mixture components sharply. The results show that the proposed algorithm can not only reduce the running time, but also guarantee the tracking performance with a proper confidence interval.