摘  要

行人骨架关节角度估计是当今机器视觉领域的研究热点，它涉及到图像处理、模式估计、人工智能等多个学科的相关知识。基于2D连续图像序列的行人骨架关节角度估计包括人体区域检测、动作及姿态分割以及用于进行动作估计的目标分析和行为理解等。在分析总结该领域研究人员相关工作的基础上，本文针对俯卧撑运动的特点进行动作和姿态分割，并对其完整动作进行估计。

本文的主要研究内容如下：

1、本文针对目标动作的特点提出了一种基于动作变化率特征的动作及姿态分割方法。首先获取图像中动作区域的轮廓信息，根据连续图像序列中轮廓信息的变化情况挖掘出动作的变化率，然后利用量化后的动作变化率界定动作及姿态的分割点，最后按照对动作估计的意义大小，将不同的姿态划分为关键姿态和非关键姿态。由于关键姿态携带了进行动作估计的绝大部分信息，因而只利用关键姿态进行动作估计，这种方式有效地降低了计算复杂度，提高了实时性。

2、本文针对2D连续图像序列采集和处理过程中产生的偶然性误差提出了一种基于向量模的误差消除算法，用于消除数据序列中异常数据引起的误差，其基本原理是在原有数据序列的基础上利用多维向量的模构建一组新的数据序列。首先采用某一数据节点及其相邻数据模拟多维向量各方向上的各个分向量，然后计算此多维向量的模并将其作为与当前数据节点对应的新数据序列中的节点。

3、本文在动作及姿态分割的基础上采用了基于棍状模型的动作估计方法，通过分别建立各关键姿态的人体模型并与实际动作人体进行比较的方式进行目标分类和动作估计。结合大量的实验对以上方法和算法进行稳定性和准确性的验证，验证结果令人满意。

关键词  俯卧撑  动作分割  姿态分割  向量模  模型匹配  动作估计

Abstract

The angle estimation of pedestrian skeleton joint is a hot topic in the field of machine vision. It involves many subjects such as image processing, pattern estimation, artificial intelligence and so on. The angle estimation of pedestrian skeleton joints based on video includes human region detection, motion and posture segmentation, target analysis and behavior understanding for motion estimation. On the basis of analyzing and summarizing the related work of researchers in this field, this paper divides the motion and posture according to the characteristics of push-ups, and estimates its complete motion.

The main contents of this paper are as follows:

1. According to the characteristics of target action, this paper proposes an action and posture segmentation method based on the feature of motion change rate. Firstly, the contour information of the action region in the image is obtained, and then the changing rate of the action is mined according to the variation of the contour information in the sequence of continuous images, and then the segmentation points of the action and posture are defined by the quantized change rate of the action. Finally, according to the significance of motion estimation, the different attitude is divided into critical attitude and non-critical attitude. Because the key attitude carries most of the information of motion estimation, only the key attitude is used to estimate the action. This method can effectively reduce the computational complexity. High real-time performance.

2. In this paper, an error elimination algorithm based on vector mode is proposed to eliminate the error caused by abnormal data in the data sequence. The basic principle is to construct a new set of data sequences by using the modules of multidimensional vectors on the basis of the original data sequences. First, a data node and its adjacent data are used to simulate the upward sub-vectors of the multidimensional vector, and then the modules of the multidimensional vector are calculated and used as the nodes in the new data sequence corresponding to the current data node.

3. On the basis of motion and posture segmentation, the method of motion estimation based on stick model is adopted in this paper, and the target classification and motion estimation are carried out by establishing the human body models of each key posture separately and comparing with the actual action human body. Combined with a large number of experiments, the stability and accuracy of the above methods and algorithms are verified, and the results are satisfactory.

Keywords:push-ups segmentation attitude segmentation vector module model matching motion estimation