Abstract: Constructing the appearance model of object is a key problem that affects visual tracking performance. To solve this problem, we propose an online learning discriminative model for visual tracking;a robust tracking algorithm with this model with Bayesian estimation. Firstly, we segmented the initial tracking area to generate training samples, and obtained a discriminative model via clustering of the training samples. Then, we computed a likelihood map of the predicted tracking area of the current frame using the discriminative model. Finally, we estimated the object state via maximum a posterior estimation and updated the discriminative model online. The proposed algorithm updates the appearance model via online learning, which improves adaptability for large variations in appearance. Experimental results indicate that the proposed algorithm can cope well with complex change of object's appearance, demonstrating an especially robust performance when tracking an object undergoing scaling, illumination change, occlusion, and non-rigid deformation. Both qualitative and quantitative comparisons show the superiority of the proposed algorithm to now current state-of-the-art approaches.