摘要
针对当前动作识别技术中正确识别率不高,易受到环境变化的影响等问题,提出了一种基于受限玻尔兹曼机与密集采样特征迭代加权融合的动作识别算法。避免单个特征对图像序列的表达力不强,引入了受限玻尔兹曼机（RBM）特征与密集采样（DT）特征分别对行为动作进行特征提取,得到RBM特征和DT特征;定义一种迭代加权函数,将RBM特征与DT特征进行加权融合,形成描述能力更强的RBM-DT特征;基于K-近邻（KNN）算法,对RBM-DT特征进行分类学习,完成动作识别的决策判断。通过在KTH、Hollywood数据集上实验表明：与当前动作识别技术比较,提出的新算法能够有效识别各种行为动作,对各类型动作均具有更高的正确识别率与鲁棒性。
In order to solve the defects as low correct recognition rate in the current action recognition technology,and it is easy to be affected by the change of the environment,the movement identification algorithm based on restricted Boltzmann machine and dense trajectory feature iterative weighting is proposed.The restricted boltzmann machine and the dense sampling features are introduced to extract behavior respectively for avoiding the poor expression ability of single feature to image sequence.An iterative weighting function is defined to fuse the RBM features with the DT features for forming a more powerful RBM-DT feature.The classification of RBM-DT features is classified based on the K-nearest neighbor to finish the decision judgment of the action recognition.The experiments on common KTH and Hollywood data sets show that the new algorithm can effectively identify a variety of behavior which has a high correct recognition rate and good robustness compared with the commonly used motion recognition technology.
作者
潘强
印鉴
PAN Qiang;YIN Jian(School of Economics and Management, Zhuhai City Polytechnic, Zhuhai 519090, China;School of Data and Computer Science, Sun Yat-Sen University, Guangzhou 510006, China)
出处
《光学技术》
CAS
CSCD
北大核心
2018年第2期164-170,共7页
Optical Technique
基金
国家自然科学基金(61033010)
广东省自然科学基金(S011020001182)
关键词
动作识别
受限玻尔兹曼机
密集采样
迭代加权
融合特征
K-近邻
movement identification
restricted boltzmann machine
dense trajectory
iterative weighting
fusion feature
k-nearest neighbor
