Abstract: In order to improve the accuracy of key frame extraction, and improve the quality of video summaries, a video summarization key frame extraction method for HEVC compressed domain is proposed. Firstly, the video sequence is coded and decoded, and the number of luminance prediction modes of the HEVC intra-coded PU block is counted in the decoding. Secondly, the feature extraction is constructed by using the number of patterns obtained by statistics as a pattern feature vector and used as a texture feature of the video frame for key frame extraction. Finally, the pattern feature vector is clustered by adaptive clustering algorithm, which the fusion iterative self-organizing data analysis algorithm (ISODATA). The frames corresponding to the intermediate vector in each class is selected as the candidate key frames in the clustering result, and the candidate key frames are again filtered by the similarity, which the redundant frames are eliminated to obtain the final key frames. The experimental results show that a large number of experiments on the Open Video Project dataset indicate that the precision of the key frames extraction is 79.9%, the recall rate is 93.6%, and the F-score is 86.2%, which effectively improves the quality of the video summarization.