Abstract: Traditional tracking methods utilizing the single view feature to describe samples are inaccuracy and insufficiency.Therefore,these methods will result in some problems,for instance,the noise sample as the target object usually products the drifting in the tracking processing.For above problems,we propose a robust tracking model based on multi-view features fusion.Firstly,the proposed model achieves the minimum of the training error sum of all single features by iteration solving,and introduces combinatorial coefficients with higher powers to avoid degradation of the model.Moreover,we use the kernel function to fusion multi-view features of different dimensions.Secondly,in order to achieve the real-time building of the discrimination,we improve the model to the incremental method,and obtain the candidate set of target samples at the current frame.Finally,according to the position relationship between the candidate set and the decision boundary we obtain the most suitable sample as the target object,and the exponential function is used to strengthen the differentiation between the samples.According to the tracking performance in 20 image sequences with challenging,the proposed method achieves more effective and robust performance than popular tracking methods.