Abstract: In the course of behavior identification, extracting keyframes from the video can
effectively reduce the amount of video index data, so as to improve the accuracy and real-time
performance of behavior identification. A method for optimizing the keyframe sequence is proposed
to improve the representativeness of keyframes, on which the behavior identification is based. Firstly,
the K-means clustering algorithm is employed to extract keyframes in the human action video
sequence according to 3D human skeleton features. Then, the quadratic optimization is performed in
the light of the location of keyframes to extract the optimal keyframe, and it can reduce the redundancy
of keyframe sequence, compared with traditional ways. Finally, the behavior video is identified by
convolutional neural network (CNN) classifiers in accordance with the optimal keyframe. The
experiment results on the Florence 3D Action dataset indicate that the method has a high identification
rate, and drastically shortens the identification time, compared with the traditional method.

Key words: behavior identification, keyframes, K-means, convolutional neural network