To improve the motion expression and compression rate of the motion video key frames, a dynamic video frame extraction technique based on flexible pose estimation and spatial-temporal feature embedding was proposed. Firstly, a Spatial-Temporal feature embedded Flexible Mixture-of-Parts articulated human model (ST-FMP) was designed by preserving the spatial-temporal features of body parts, and the N-best algorithm was adopted with spatial-temporal locally preserving of uncertain body parts to estimate the body configuration in a single frame based on ST-FMP. Then, the relative position and motion direction of the human body were used to describe the characteristics of the human body motion. The Laplacian scoring algorithm was used to implement dimensionality reduction to obtain the discriminant human motion feature vector with local topological structure. Finally, the ISODATA (Iterative Self-Organizing Data Analysis Technique) algorithm was used to dynamically determine the key frames. In the key frame extraction experiment on aerobics video, compared to articulated human model with Flexible Mixture-of-Parts (FMP) and motion block, the accuracy of uncertain body parts by using ST-FMP was 15 percentage points higher than that by using FMP, achieved 81%, which was higher than that by using Key Frames Extraction based on prior knowledge (KFE) and key frame extraction based on motion blocks. The experimental results on key frame extraction for calisthenics video show that the proposed approach is sensitive to motion feature selection and human pose configuration, and it can be used for sports video annotation.