摘要
人体是非刚性物体,随着人体形状的变化,底层特征也会发生剧烈的改变,如何在剧烈变化的人体动作中发现其不变性是解决人体动作识别问题的关键。提出了一种基于层级化特征的人体动作识别算法。首先,基于Harris3D检测算法从动作视频中提取人体动作的底层特征;之后,根据人体结构对特征进行区域性的划分,并根据区域划分的精细程度得到高,中,低三层特征集合。利用词袋模型对特征点进行统计,使不同层级的特征映射到相同的特征维度空间。最后,使用隐条件随机模型对人体动作进行训练和识别。大量的实验结果也证明了层级化特征在人体动作识别上的稳定性。
Since the human body has the non-rigid structure, the low-level feature will change sharply along with the change of human shape. It is a key point for human action recognition to find the invariability from human action changes. This article proposes a human action recognition algorithm based on hierarchical feature. First, the low-level featuresare extracted from each frame of a video which contains human action using the Harris 3D detection method, and is classified into different parts of human body according to the skeleton position information. Then, the hierarchical feature is computed by the definition of different layer. The bag of words model is utilized to map features of different layers into the same dimensional space. Finally, the hidden conditional random field classifier is used to model the visual transition of a human action sequence. The experiments demonstrate that the hierarchical feature has well reliability and can efficiently promote the accuracy of human action recognition.
出处
《信息技术》
2015年第11期147-151,共5页
Information Technology
关键词
层级化
视频底层特征
人体动作识别
Hierarchical
video low-level feature
human action recognition
