Real-time keyframe extraction is significant for the on-line video retrieval and recognition system. It will lead to low speed for keyframe extraction by analyzing the frames one by one on the basis of shots detection. A novel keyframes extraction method based on dynamic time-domain detection windows is presented. The video frame-sequence is divided into video units with different lengths. The frames at the edges of the window and the sampled frames inside the window are analyzed to judge the changes of the unit content. Meanwhile, a hybrid growth mode of windows and slow detection mechanism are proposed to ensure the quality and speed of the keyframe extraction. The extracted keyframes are clustered to obtain the best keyframe. The experiments show that the proposed method has a good comprehensive performance for different kinds of videos.