基于时频域特征的人员连续动作分段识别

基于惯性传感器的步态识别研究是人工智能应用到实际生活的典型范例,近几年取得满意的成就.针对日常生活中连续动作类型的信号,对其进行精确分割和识别的效果仍略有不足;这些研究局限于传感器信号的时域特征和一些简单的频域特征,且没有对不同动静状态的动作进行分类.将人类常见6种连续行为分为3类动态动作和3类静态动作,并对其进行分割和识别.使用滤波器去除原始信号噪声干扰.通过滑动窗口分割法进行分割,对每一个窗口片段提取常用传统特征和梅尔倒谱系数,以及倒谱系数的一二阶delta导数等频域特征.将6类动作的不同特征进行多种组合,使用分类器识别不同动作,将不同特征组合的识别结果进行比较.该模型在UCI公开数据集随机抽取了5组测试样本,整体分割识别准确率最高达到98.19％.

Segmental recognition of human continuous motion based on time-frequency domain features

Research on gait recognition based on inertial sensor is a typical example of artificial intelligence applied to real life and has achieved satisfactory achievements in recent years.However, for the signals of continuous action in daily life, the effect of accurate segmentation and recognition is still a little inadequate.Secondly, these studies are limited to the time-domain characteristics and some simple frequency-domain characteristics of sensor signals, and there is no classification of actions in different static and static states.In this paper, six kinds of human continuous behaviors are divided into three kinds of dynamic actions and three kinds of static actions, and they are segmented and recognized.First, a filter is used to remove the noise from the original signal.The sliding window segmentation method is used to extract the traditional features and Meir cepstrum coefficients, as well as the first and second order delta derivatives of the cepstrum coefficients in frequency domain.Finally, the different features of the six types of actions are combined in a variety of ways, and the classifier is used to identify different actions, and the recognition results of different feature combinations are compared.The model randomly selected five groups of test samples from UCI public data set, and the overall segmentation and recognition accuracy reached 98.19%.