摘要
提出基于关键帧技术的单次学习的动作识别方法。首先把样本动作建模为若干关键姿势的集合,然后利用关键帧把测试视频分割为单个动作,最后结合关键姿势相似度的时间加权计算动作匹配概率,进行动作识别。该动作识别方法只通过单次学习,提高了效率。使用标准深度数据库进行实验,实验结果显示了该方法的有效性,尤其在区分差别比较细微的动作方面。
As human can distinguish two different actions according to some gestures,a key-frame-based one shot learning action recognition method was proposed.Firstly,the sample actions were modeled by sets of several key frames.Then using key frames,the test videos were segmented to videos which include only one action.Next,considering the time weighted similarity of key frames,the matching probability between two actions was calculated.As it is a one-shot learning action recognition method,the efficiency is improved.Experiments were carried out with standard depth data,and the results prove the effectiveness of the proposed method,especially for classifying actions with little difference.
出处
《半导体光电》
CAS
北大核心
2015年第6期999-1005,共7页
Semiconductor Optoelectronics
基金
上海市自然科学基金项目(14ZR1446100)
关键词
关键姿势
动作识别
单次学习
深度图像
视频分割
key frame
action recognition
one shot learning
depth image
video segmentation