面向行为识别的拉普拉斯特征映射算法的改进

Improvement of Laplacian eigenmaps for human action recognition

提出了一种面向行为识别的拉普拉斯特征映射算法的改进方法。首先,将Kinect提供的关节点数据作为姿态特征,采用Levenstein距离改进流形学习算法中的拉普拉斯特征映射算法,并映射到二维空间得到待识别行为的嵌入空间;其次,结合待识别行为的嵌入空间和训练数据建立先验模型;最后,通过重新设计的粒子动态模型和观察模型,采用粒子滤波算法进行行为识别。实验结果表明,该方法可以对重复动作、遮挡,以及动作幅度和速度都有明显差异的行为进行较好的识别,总体识别率达到92.4%。

This paper presented a method for human action recognition by improving Laplacian eigenmaps algorithm. Firstly,Kinect sensor offered joints data as posture feature. Laplacian eigenmaps algorithm which one of the manifold learning was improved by Levenstein distance mapped the features to two-dimensional space which was the embedding space for behavior recognition. Secondly,the method built a prior model combined embedding space and training data. Finally,by redesigning particle dynamic model and particle observation model,the method employed particle filter algorithm to recognize behavior. The experimental results show that the proposed method can obtain satisfactory results among behaviors existed repetitive actions,shield movement and visible difference in range of movement and speed. The recognition rate is 92. 4%.