Multitarget tracking is very important in intelligent video surveillance system. Occlusions among many moving objects and objects with similar appearance that could result in incontinuous trajectory are challenging problems in this field. Based on the fact that the relative positions remain stable between two close pairs, a novel multitarget tracking algorithm based on conditional random field is presented in this paper. Unlike previous approaches which focus only on appearance and motion models for all targets, we consider discriminative features for distinguishing difficult pairs of targets with sets of labels. Multilevel is adopted by this means, in which the set of tracklets produced by previous level is used as an input. And tracking problem is transformed into an energy minimization problem including a set of unary function for a continuous trajectory and a set of pairwise function for pairs of tracklets. This new method is more powerful in deal with objects with similar appearance and badly occlusions compared to stateofart methods. Qualitative and quantitative experimental results show that this new method has a better performance.