For each video sequence of different resolutions and different QP, a SVM classifier is pre-trained respectively.The training Video sequences of different resolutions are obtained from 4K video down-sampling. In order to maximize the prediction accuracy, each SVM classifier would be updated online according to the information getting from the first frame in a group of intra pictures (GOIP). However, in some cases, the number of CU getting in the first frame may be small, and thus there is no sufficient information to update the SVM classifier. In this case, the off-line trained SVM would be directly used for the CU early termination decision process in current GOIP.