基于双目视觉的人体行为分析技术研究

发布时间：2018-12-18 18:34

【摘要】：人体行为分析技术是计算机视觉领域的一个研究热点问题。该技术在视频监控、感知接口、运动分析和虚拟现实等多个领域均具有广阔的应用前景。其中如何有效克服遮挡和多义性、环境的复杂变化性以及人体的非刚体性等困难的影响成为人体行为分析技术中的一个重要任务。基于此,本文围绕基于双目视觉的人体行为分析技术展开研究,重点针对基于双目视觉的立体匹配与深度信息获取方法和基于卷积神经网络的人体行为分析算法展开了分析与研究,提出了一些解决方法和改进措施。本文研究的主要内容如下:1、在基于双目视觉的立体匹配与深度信息获取算法研究中,提出了一种基于人体边缘信息的SURF(Speeded-Up Robust Features-简称SURF)与区域匹配结合的立体匹配算法。该算法旨在降低遮挡和多义性造成的影响,引入三维深度信息提高行为分析算法的精度。该方法包括双目视觉系统标定、运动目标检测、SURF立体匹配与区域匹配优化、三维信息获取四个部分。在采用平面模板两步法完成双目视觉系统的标定后,采用改进的混合高斯模型的背景差分法提取人体运动目标。在匹配过程中,先对获取的人体边缘信息进行SURF匹配,然后结合基于极限约束的区域匹配算法进一步优化匹配结果,提高人体特征点匹配的精度。最后根据得到的匹配点获取三维深度信息。实验结果表明,该算法能够准确获取人体三维空间坐标,有效避免遮挡和多义性的干扰。2、在基于双目视觉的人体行为分析算法研究中,提出了一种基于小样本卷积神经网络(Convolutional Neural Networks-简称CNN)的人体行为分析算法。卷积神经网络分为特征提取层和特征映射层。在特征提取层,利用CNN神经元感知并提取局部特征;然后利用由多个特征映射层组成的网络层进行相应的计算,使得特征提取精度更为准确可靠。基于小样本卷积神经网络的人体行为分析算法分别对双目视觉系统下左右相机采集的图像采用CNN方法进行分类识别,然后对左右图像的识别结果进行权值融合处理,通过调节系统参数,获取更高的行为匹配度。实验结果表明,该算法能够对单人动作和交互动作进行准确识别,有效提高人体行为分析算法的识别率。
[Abstract]:Human behavior analysis is a hot topic in the field of computer vision. This technology has broad application prospects in many fields such as video surveillance, perceptual interface, motion analysis and virtual reality. How to effectively overcome the influence of occlusion and polysemy, the complexity of environment and the non-rigid nature of human body has become an important task in human behavior analysis technology. Based on this, this paper focuses on the research of human behavior analysis technology based on binocular vision. The methods of stereo matching and depth information acquisition based on binocular vision and the algorithm of human behavior analysis based on convolutional neural network are analyzed and studied, and some solutions and improvement measures are put forward. The main contents of this paper are as follows: 1. In the research of stereo matching and depth information acquisition algorithm based on binocular vision, A stereo matching algorithm combining SURF (Speeded-Up Robust Features- SURF) and region matching based on human edge information is proposed. The algorithm aims to reduce the influence of occlusion and polysemy and improve the accuracy of behavior analysis algorithm by introducing 3D depth information. The method includes four parts: binocular vision system calibration, moving target detection, SURF stereo matching and region matching optimization, and 3D information acquisition. After the calibration of the binocular vision system was completed by using the plane template two-step method, the background difference method of the improved mixed Gao Si model was used to extract the moving target of human body. In the process of matching, the human body edge information is first matched by SURF, and then the matching result is optimized by combining the region matching algorithm based on limit constraint to improve the accuracy of human body feature point matching. Finally, the 3D depth information is obtained according to the matching points. The experimental results show that the algorithm can accurately obtain the three-dimensional coordinates of human body and avoid the interference of occlusion and polysemy. 2. In the research of human behavior analysis algorithm based on binocular vision, A human behavior analysis algorithm based on small sample convolution neural network (Convolutional Neural Networks- for short CNN) is proposed. Convolution neural network is divided into feature extraction layer and feature mapping layer. In the feature extraction layer, the CNN neuron is used to perceive and extract the local features, and then the network layer composed of multiple feature mapping layers is used to calculate the feature extraction accuracy more accurately and reliably. The human behavior analysis algorithm based on small sample convolution neural network uses CNN method to classify and recognize the images collected by left and right cameras in binocular vision system, and then carries on the weight fusion processing to the recognition results of left and right images. By adjusting the system parameters, a higher behavior matching degree can be obtained. The experimental results show that the algorithm can accurately identify single action and interactive action, and improve the recognition rate of human body behavior analysis algorithm.
【学位授予单位】：北方工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】