静态图像中的人体行为分类研究  

Study on human action classification in still images

作　　者：马秀梅[1] 唐春晖[1] 尹征[1] 张仁杰[1] 

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《信息技术》2015年第4期98-101,共4页Information Technology

摘　　要：静态图像中人体行为分类的一般方法是先手动标定出行为对象,再单独对行为对象进行特征提取和分析,不仅费时费力还丢失了场景信息。针对此问题提出了结合场景特征与行为对象特征的图像表示方法,充分利用图像的所有信息。此外为了减小量化误差,在特征编码阶段,采用局部约束线性编码(Locality-constrained Linear Coding,LLC)算法,并将其应用在空间金字塔模型的向量量化中。该算法与传统的矢量量化算法和稀疏编码算法相比,能够降低量化误差。最后在Stanford 40 Action数据集上对文中方法进行实验和验证,结果表明,结合场景特征与行为对象特征并使用LLC编码算法能够获得更好的分类效果。The general method of human action classification in still images is to calibrate the action object manually,and then extract the features of the separated action object from the original image to perform an analysis solely,it is time-consuming and laborious by this way,in addition,the object scene context is largely suppressed. For this problem this paper combined the scene and action object for utilizing scene information. Furthermore,in order to reduce the quantization error,it utilizes localityconstrained linear coding（ LLC） algorithm in the coding phase and apply LLC to vector quantization of spatial pyramid matching model,compared with the traditional vector quantization and sparse coding,LLC can improve classification result. Finally it makes experiment in Stanford 40 Actions dataset with the method. The results show that combined with scene feature and action object feature it can get the better classification result by using LLC algorithm.

关 键 词：行为分类 局部约束线性编码 特征提取 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]