基于K-SVD编码和金字塔词汇森林的行为识别

摘要

词袋模型在构造视觉字典时,对特征做单一划分,容易造成误差,且其忽略了局部特征在时空中的关系,为此提出一种基于K-SVD编码和时空金字塔词汇森林的识别方法.使用K-SVD稀疏编码的方式构造字典,将特征划分到多个类别中,通过多个视觉单词的线性加权来表示特征向量,减少单一划分的误差,增强分类能力,通过构造时空金字塔词汇森林对特征的结构信息做进一步描述,获得更加丰富且具有区分度的分类模型.实验结果表明,该方法进一步描述了特征中的潜在信息,行为识别精度高达97.33%.%When using the bag of words model to construct visual dictionary, the feature is partitioned singly, causing the error, and the relationship of local features in time and space are neglected.To solve the problem, a recognition method based on K-SVD coding and space time pyramid vocabulary was proposed.The dictionary was constructed using K-SVD sparse coding me-thod and the feature was divided into multiple categories.The feature vector was represented by linearly weighting multiple vi-sual words to reduce the error of single division and enhance the classification ability.The structure information of feature was described through constructing temporal and spatial pyramid vocabulary.Experimental results show that the proposed method further describes the characteristics of the potential information, and the behavior recognition accuracy is improved to 97.33%.