融合形状和运动特征的动作识别计算模型  

Computational Model Combining Form and Motion Features for Human Action Recognition

作　　者：刘海华[1] 郝丽芳[1] 谌先敢[1] 

机构地区：[1]中南民族大学生物医学工程学院,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2013年第2期81-86,共6页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家自然科学基金资助项目(60972158);湖北省自然科学基金重点资助项目(2011CDA078);武汉市科技计划资助项目(201271130454)

摘　　要：针对视觉系统在动作识别过程中如何利用形状与运动信息的问题,提出了一种融合形状特征和运动特征的人体动作识别方法.该方法模拟视觉皮层的背侧和腹侧通路,建立了基于双通道理论的人体动作特征计算模型.计算模型分别利用2D Gabor滤波器和3D时空滤波器模拟腹侧和背侧通路中视觉皮层简单细胞,提取动作的时空信息,通过采样、局部遍历、模板学习一系列操作分别提取动作的时空特征,并采用线性融合方法获取描述动作的特征向量,构建了采用支持向量机(SVM)进行动作分类的动作识别系统.实验结果表明:该方法的识别性能优于同类型的识别方法,取得了较好的识别效果.This paper presents a computational model for human action recognition, which combined the form and motion features. This method mainly simulates the visual system~ workflow about how to take use of form and motion information during the process of motion recognition. It established the computational model based on the theory of two channels by simulating the dorsal and ventral streams. In order to get the form and motion information,2D gabor filters and 3D spatial time filters are used separately to simulate the simple ceils in the visual cortex. Then the same processing pipeline is applied in both channels： feature maps are pooled locally, down-sampled, and compared with a set of learnt templates, yielding a vector of similarity scores. In the final step, the two score vectors are merged and the final feature vector describing the human action will be achieved. It established the action recognition systems for action classification with support vector machine （SVM）. The result shows that our method outperforms the similar existing methods, and gets better recognition.

关 键 词：动作识别 特征融合 特征提取 支持向量机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]