新的时空特征点检测方法
尹建芹;王晶晶;李金屏
【期刊名称】《吉林大学学报（工学版）》 【年(卷),期】2012(042)003
【摘要】A new scheme to detect space-time interest points based on cumulative entropy difference is proposed for human action recognition and
video
analysis.First,periodic
space-time
features
are
detected.Second,the concept of cumulative entropy difference is put forward as the criterion to evaluate the interest points.Third,the points with high cumulative entropy are selected as the final key points.Finally,c-means cluster analysis is applied to find the features,and the prototype of the key video is obtained.The video is converted to the feature vector presented by the prototype;then the video classification is conducted.Experimental results show that the cumulative entropy difference evaluation can effectively remove the noise cuboids,and the method based on the evaluation can realize human action analysis and facial expression recognition.%提出了基于累积差熵的时空特征点检测算法。首先利用周期性时空检测方法检测视频的关键特征点;然后提出了视频累积差熵的概念,用累积差熵作为特征点的评价准则;以该准则为基础,选择具有累积差熵大的特征点作为关键点,并对关键视频进行聚类,得到关键视频的原型特征。实验结果表明：本文方法可以简单有效地去除非运动信息得到的关键点,可以较好地用于动作识别、表情识别等视频分析领域。
新的时空特征点检测方法
