视频中人体行为识别的研究

发布时间：2019-05-31 15:12

【摘要】：视频中的人体行为识别技术一直以来都是计算机视觉领域中一个非常活跃研究课题,在智能监控、视频检索、人机交互等领域应用广泛。本文主要研究视频中人体行为识别的方法。主要思路是通过时空兴趣点获得极具价值的行为底层特征信息,通过视觉词袋模型得到行为特征的人体行为知识词典,通过行为特征训练学习产生人体行为模式。为了提高复杂场景下的人体行为识别的准确率,本文对传统的特征提取和描述方法加以改进,主要的工作内容如下:(1)深入探讨了人体行为识别过程中各个阶段的关键技术。对3D Harris、Dollar时空兴趣点检测方法做了研究和分析,探讨了时空兴趣点特征描述方法,包括HOG、HOF、3D HOG描述子。运用视觉词袋模型进行行为类别建模,最后,结合SVM分类器实现对视频中人体行为的判定。仿真结果表明,这些方法在KTH和UCF Sports数据库上有很好的识别效果。(2)针对复杂背景中会产生一些无关兴趣点的问题,给出了增加时空兴趣点质量的方法。通过对人体运动区域的提取,限定时空兴趣点的产生范围,可以抑制复杂背景带来的不利影响,来增加提取时空兴趣点的质量。通过仿真发现该方法在去除干扰时空兴趣点的同时,对于识别效果的提高也是有效的。(3)从丰富特征描述的信息出发,提出了使用复合时空特征和融入时空兴趣点的轨迹特征对时空兴趣点及邻域进行描述。复合时空特征结构中包含了局部的空间结构信息和局部的运动信息,而进一步提出的融入轨迹特征的描述子还具备了局部的相对运动信息。从实验的结果可以看出,这两个种新的描述方法对于提高行为识别的效果都有好处,融入时空兴趣点的轨迹特征比复合时空特征表现更好。
[Abstract]:Human behavior recognition technology in video has always been a very active research topic in the field of computer vision, and has been widely used in intelligent monitoring, video retrieval, human-computer interaction and other fields. This paper mainly studies the methods of human behavior recognition in video. The main idea is to obtain valuable behavior characteristic information through space-time interest points, to obtain human behavior knowledge dictionary of behavior characteristics through visual word bag model, and to produce human behavior patterns through behavior feature training and learning. In order to improve the accuracy of human behavior recognition in complex scenes, this paper improves the traditional feature extraction and description methods. The main contents are as follows: (1) the key technologies in each stage of human behavior recognition are discussed in depth. The detection method of space-time interest points in 3D Harris,Dollar is studied and analyzed, and the feature description method of space-time interest points, including HOG,HOF,3D HOG descriptors, is discussed. The visual word bag model is used to model the behavior category. Finally, the SVM classifier is combined to determine the human behavior in the video. The simulation results show that these methods have a good recognition effect on KTH and UCF Sports databases. (2) aiming at the problem that some independent interest points will occur in complex background, a method to increase the quality of space-time interest points is given. By extracting the motion region of human body and limiting the generation range of space-time interest points, the adverse effects of complex background can be suppressed, and the quality of space-time interest points can be increased. Through simulation, it is found that this method is also effective for improving the recognition effect while removing the interest points of interference space-time. (3) starting from the rich information of feature description, In this paper, the compound space-time characteristics and the trajectory characteristics of space-time interest points are proposed to describe the space-time interest points and neighborhoods. The compound space-time feature structure contains the local spatial structure information and the local motion information, and the further description which integrates the trajectory feature also has the local relative motion information. From the experimental results, it can be seen that these two new description methods are beneficial to improve the effect of behavior recognition, and the trajectory characteristics of integrating space-time interest points are better than those of compound space-time features.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】