采用自适应变异粒子群优化SVM的行为识别  

Action recognition based on adaptive mutation particle swarm optimization for SVM

作　　者：张国梁[1] 贾松敏[1] 张祥银[1] 徐涛[1] 

机构地区：[1]北京工业大学信息学部,北京100124

出　　处：《光学精密工程》2017年第6期1669-1678,共10页Optics and Precision Engineering

基　　金：国家自然科学基金项目(No.61175087);北京工业大学智能机器人"大科研"推进计划"助老智能轮椅床自主测控系统的研究与实现"资助项目(No.040000546317552)

摘　　要：为了提高对视频序列中人体行为的识别能力,建立了基于局部特征的动作识别框架。通过时空特征提取及编码和SVM分类器参数优化两部分对该框架所涉及算法进行了研究。首先,采用Harris3D检测器获取时空兴趣点(STIP),以方向梯度直方图(HOG)和光流方向直方图(HOF)对STIP进行描述,并引入Fisher向量实现对特征描述子的编码;由于固定参数下SVM动作分类模型存在泛化能力不足的问题,将粒子群算法应用于各动作分类器参数寻优过程中,针对种群多样性逐代变化的特点,构建粒子聚集度模型,并利用其动态调节各代粒子的变异概率;最后,利用KTH和HMDB51数据集对所提方法进行验证。结果表明,所提自适应变异粒子群算法(AMPSO)能够有效避免种群陷入局部最优,具备较强的全局寻优能力;在KTH和HMDB51数据集上的识别准确率分别为87.50%和26.41%,优于其余2种识别方法。实验证明,AMPSO算法收敛性能良好且整体识别框架具有较高的实用性和准确性。The action recognition framework based on local features was established to improve the recognition ability of human behavior in video sequences.The algorithms related to the framework were researched through spatial temporal features extracting and encoding and parameters optimization of SVM classifier.Firstly,the feature descriptors composed of Histograms of Oriented Gradients（HOG）and Histograms of Optical Flow（HOF）were used to describe Space Time Interest Points（STIP）achieved by the Harris 3Ddetector and then encoded by Fisher Vector（FV）.Due to the generalization ability of Support Vector Machine（SVM）model for action classification under fixed parameters was insufficient,the particle swarm optimization algorithm was applied to the parameter optimization of each action classifier.According to the characteristics of population diversity changed from generation to generation,the constructed particles aggregation degree model was used to adjust mutation probability for each generation of particles dynamically.Finally,the proposed method wasverified by KTH and HMDB51 data sets.The results show that the Adaptive Mutation Particle Swarm Optimization（AMPSO）algorithm can avoid the local optimum and has strong global optimization capability.The recognition accuracies on KTH and HMDB51 data sets are 87.50%and 26.41%,respectively,which are better than two recognition methods.Experimental results indicate that the AMPSO algorithm has good convergence performance and the overall recognition framework has high practicability and accuracy.

关 键 词：人体行为识别 自适应变异粒子群算法 时空兴趣点 特征编码 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]