In view of the problem that the moving target may be occluded or tracked under uncertain conditions during the aerial target tracking of UAV in complex scenes, resulting in the gradual damage, drift, and irreversible failure of the visual model, a long-term tracking algorithm for UAV was proposed. Firstly, the complementary classifier was designed for multi-feature adaptive fusion. The color histogram feature was used in Bayesian classifier, and the directional gradient histogram, grayscale, and color name features were used in the correlation filter. In combination with the advantages of multiple features, the robust appearance of the target was built to adapt to the complex scenes. Secondly, an adaptive spatial-temporal regularization term was added to the correlation filter. Local changes were introduced into the spatial regularization parameters for the implementation of the filter with low pixel credibility during learning. In temporal regularization, the learning of the filter was adaptively adjusted according to the global response, and the initial filter was used to constrain the update range, which effectively prevented filter degradation while mitigating boundary effects. Finally, a re-detection module was added to make sure the accuracy of the tracking process. Experimental results show that the proposed algorithm could adapt to the complex scenes of UAV aerial photography, alleviate boundary effects, and prevent filter degradation. In comparison with similar mainstream algorithms, the proposed algorithm could still meet the real-time requirements and achieve better tracking effect even when the target experienced serious occlusion or moved out of view.