An efficient approach based on wavelet transform is presented to detect the variation degree of dynamic background. Firstly, some candidate regions can be obtained by processing IR data through wavelet transform, then the degree of the difference between current and referenced background can be judged through detecting the difference of the above feature regions. Finally, some small moving candidate targets can be detected by eliminating background. After eliminating background, data association and robust tracking of those targets can be realized by level set tracking method based on region-based active contour with dynamic priors. Dynamic priors include shape descriptor, intensity information, moving features and etc. In addition, targets’ merging and splitting phenomena occasionally arise during multi-target tracking. This problem can be solved by ‘memorizing and filling’. The proposed approach is validated to track multi-target effectively by using actual infrared image sequences with complex dynamic background. Experiment results indicate the feasibility and effectiveness of the proposed method.