基于关键点亲和场与支持向量机的人体姿态识别

HUMAN POSTURE RECOGNITION BASED ON POINT AFFINITY FIELD AND SVM

针对传统动作姿态识别仍需物理数据采集设备或深度体感设备进行手工提取特征的问题,提出一种基于关键点亲和场与支持向量机的人体姿态识别方法。以关键点亲和场为核心进行关节点检测,获取各种姿态下的18个关节点坐标信息,使用标准化后的坐标数据对支持向量机模型加以训练,选择不同的高斯核函数进行对比。实现在没有人体深度信息和无穿戴设备的情况下,只使用普通RGB图片便可对人体姿态进行分类识别的效果。实验表明它在KTH数据集、Weizmann数据集中的识别效果良好;在自采集数据集中与带有传感器的方法相比,缩减操作步骤的同时准确率提高了7百分点。另外,还在保持关节点检测不变的情况下,使用随机森林、KNN算法进行姿态分类对比,实验结果证明该方法优于后两者。

Aimed at the problem that traditional posture recognition still needs physical data acquisition equipment or depth somatosensory equipment to manually extract features,a human posture recognition method based on point affinity field and support vector machine is proposed.The point affinity field was taken as the core to conduct the joint detection,and the coordinate information of 18 joints of various postures was obtained.The standardized coordinate data was used to train the SVM model,and different Gaussian kernel functions were selected for comparison.In the absence of human depth information and no dressing equipment,only normal RGB images were used to categorize human postures.The experiments show that it has good recognition effect in KTH and Weizmann data sets.In the self-acquisition data set,the accuracy is improved by 7 percentage points compared with the method with sensors while the operation steps are reduced.In addition,random forest and KNN algorithm are used for posture classification comparison while keeping the detection of the joints unchanged.Experimental results show that this method is superior to the latter two.