Key frame extraction method is regarded as one of the most important i ssues in content-based video retrieve (CBVR) technology.In this paper,an efficient and stable key frame ext raction algorithm based on compressive sensing and entropy computing is proposed.Firstly a very sparse ran dom projection matrix that satisfies the condition of restricted isometry property (RIP) is constructed,whi ch is used to convert the high dimensional multi-scale frame image feature to low dimensional multi-scale fra me image feature in order to generate the column vector group of the low dimensional multi-scale feature for each video shot,and then the matching feature of each frame in one shot is calculated one by one through Hada mard product between a random weights vector and the low dimensional multi-scale feature of the frame.In the next step,the Euclidean similarity measurement between adjacent matching features is used to perform sub-shot segm entation in each shot,and finally two possible key frames are obtained in every sub-shot through cross-entropy c omputing and the ultimate key frame is selected by image entropy computing to represent the content of the sub -shot.Our experimental results demonstrate that key frames extracted by the proposed method can describe conten ts of video shots more accurately and stably than the traditional methods.