A video prediction method based on adaptive hierarchical moving kinematic modeling, comprising: establishing a hierarchical kinematic model for a video, training the model, and using the trained model to predict future frames of the video, wherein the establishment of the hierarchical kinematic model for the video comprises processes for hierarchical kinematic modeling and adaptive motion transformation for the video, and the processing for the two processes may both independently improve the effect of video prediction. Employing the method of the present invention may increase the similarity between the signal-to-noise ratio of a prediction result and the structure of an actual result and may improve the effect of video prediction, namely improving the sharpness and reasonableness of the prediction result.