一种基于多特征多分类器融合的人体行为识别方法

Human Action Recognition Method Based on Multi-feature Multi-classifier Fusion

提出了一种基于多特征多分类器融合的人体行为识别方法。针对Kinect传感器提取的三维骨骼动作序列,采用身体部位的相对几何特征、关节点的相对位置特征、关节点的绝对位置特征对人体动作进行描述。将支持向量机和随机森林分类器作为成员分类器,对3种动作特征分别进行训练和测试,使用分类器融合算法对分类结果进行融合决策,实现最终的分类。在现有的人体动作数据集上进行验证,实验结果表明:本方法可取得95%的识别率。

A human behavior recognition method based on multi-feature and multi-classifier fusion was proposed. For the three-dimensional skeleton motion sequence extracted by the Kinect sensor,the human body motion was described by using the relative geometric features of the body part,the relative position characteristics of the joint points and the absolute position characteristics of the joint points. The support vector machine and the random forest classifier were used as member classifiers to train and test the three kinds of action features respectively. The classifier fusion algorithm was used to make fusion decision for the classification results and achieve the final classification. The verification was performed on the existing human motion dataset. The experimental results show that the method can achieve 95% recognition rate.