Abstract: In order to understand the features learning process, reduce storage of data and improve recognition accuracy, an improved KNN method is proposed for human identification based on Kinect v2’s facial data and skeletal data. First of all, 3-D position information of facial feature points and skeletal joints is extracted by Kinect v2 and then the characteristic information of strong understanding like eye width and arm length can be calculated. It is proposed that an improved truncated mean method can restrain singular values by sorting data and intercepting the intermediate data, and an improved KNN method based on the accuracy of matching recognition is applied to predict human identity. Experimental results show that the proposed clustering method has higher accuracy of matching recognition and the improved classification method improves the accuracy of recognition.

Key words: human identity recognition, facial data, skeletal data, sort truncated mean method, accuracy of matching recognition

摘要： 为了理解特征学习过程、减少数据存储和提高识别率，提出使用Kinect v2的面部数据和骨骼数据作为数据集和一种改进KNN算法对人体身份的识别。使用Kinect v2提取出人体脸部特征点和骨骼关节点的三维位置信息，通过提取出的特征点的坐标计算出理解性强的特征信息如眼宽、臂长等。利用一种改进的截断均值聚类方法，通过排序把奇异值分布到数据集两端，截取数据集中间特征以抑制奇异值，利用基于匹配识别准确度的改进KNN算法对人体身份进行预测。实验结果表明提出的聚类方法匹配识别准确度更高，改进的分类方法也提高了识别的准确率。

关键词: 人体身份识别, 脸部数据, 骨骼数据, 排序截断均值法, 匹配识别准确度