采用自适应变异粒子群优化SVM的行为识别

发布时间：2018-08-23 15:45

【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。
[Abstract]:In order to improve the recognition ability of human behavior in video sequences, a motion recognition framework based on local features is established. The algorithms involved in this framework are studied in two parts: temporal and spatial feature extraction and coding and parameter optimization of SVM classifier. Firstly, the spatio-temporal interest point (STIP),) is obtained by Harris3D detector, and the STIP is described by the directional gradient histogram (HOG) and the optical flow direction histogram (HOF), and the Fisher vector is introduced to encode the feature descriptor. Due to the lack of generalization ability of SVM action classification model with fixed parameters, particle swarm optimization (PSO) algorithm is applied to the optimization of the parameters of each action classifier. According to the characteristics of population diversity generation by generation, the particle aggregation model is constructed. It is used to dynamically adjust the variation probability of each generation of particles. Finally, the proposed method is verified by using KTH and HMDB51 data sets. The results show that the proposed adaptive mutation particle swarm optimization (AMPSO) algorithm can effectively avoid population falling into local optimum and has strong global optimization ability, and the recognition accuracy on KTH and HMDB51 datasets is 87.50% and 26.41% respectively, which is superior to the other two recognition methods. Experiments show that the AMPSO algorithm has good convergence performance and high practicability and accuracy.
【作者单位】：北京工业大学信息学部;
【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)
【分类号】：TP18;TP391.41

【相似文献】