Abstract: In order to get the main information automatically, while the redundant information is less, the LLE-Adaptive FCM and LLE-Adaptive threshold FCM are presented in this paper.At the beginning of two algorithms, getting the feature vectors of video frames via locally linear embedding (LLE), and then inputing the feature vectors into LLE-Adaptive FCM and LLE-Adaptive threshold FCM, the results of clustering and Clustering centers will be got.Adaptive FCM through clustering validity function to determine the classification numbers, and Adaptive threshold FCM through the automatic changes of threshold to determine the number of classification.Finally, computing the Video Abstracts which have the shortest distances between video frames and clustering centers.The exams show that the extracting key frames can express the main content of video and redundant information is less when without artificial intervention.