基于隐马尔科夫模型的特定姿态检测

发布时间：2018-06-19 17:11

本文选题：HMM应用 + 姿态检测　；参考：《电子科技大学》2014年硕士论文

【摘要】：在高度信息化的社会,视频监控已经广泛应用于生活的每个角落。目前的监控方式还是属于传统的记录形式,对于一些重点场所还需要以人工的方式值守在监视器前。对于井喷式增长的监控数据而言,这显然不是一个高效且经济的监控方式。伴随着计算机视觉、图像处理技术的飞速发展,智能视频监控技术也逐渐成为研究领域的一大热点。在智能监控应用场景中,如何有效地识别出某些特定姿态行为的姿态具有很强的实用性和研究价值,因此已经成为智能视频监控研究的重点方向之一。由于监控场景的往往都很复杂,为了能获得很好的姿态检测结果就需要一套行之有效的姿态检测方案。在借鉴隐马尔科夫模型在语音识别领域成熟应用的基础上,本文提出了一种基于隐马尔科夫模型的姿态检测技术。该方案包括的完整检测过程包括:前景检测、前景过滤、目标跟踪、特征提取以及获取姿态序列几个主要步骤。根据人体姿态动作具有时空连续性这一特点,该方案选取了人体姿态发生过程中的时空信息作为特征,对人体姿态进行分析。在方案中采用基于时间和空间的特征分析方法,有效地刻画了人体姿态。对于人体姿态在图像两帧序列中的时间信息,本方案采用光流方法,提取人体姿态的运动信息,将获得的幅度和相位分布情况作为运动特征。对于在单帧图像的空间信息,即人体姿态在某个时刻的动作,本文基于一种快速的梯度响应图算法,提出了新的关于人体姿态匹配方法。充分利用了局部区域的关键动作代替了整体的人体姿态,简化了匹配过程中的复杂度。同时,在模板匹配的过程中,利用不同姿态和不同角度建立了层次性的模板匹配,采用该方法能在提高准确率的同时有效保证了检测速度。最后,将获得的时间和空间特征信息,基于隐马尔科夫模型获得人体姿态的状态序列,以此作为检测结果。在文中,分章节的形式介绍了本方案的技术框架,以拍照姿态检测为例实现了基于隐马尔科夫模型的拍照姿态检测原型系统。实验表明,该方法具有较强的鲁棒性和实时性。
[Abstract]:In the highly information society, video surveillance has been widely used in every corner of life. The current monitoring method is still a traditional record form, for some key places still need to be manned in front of the monitor. This is clearly not an efficient and economical way to monitor blowout growth data. With the rapid development of computer vision and image processing technology, intelligent video surveillance technology has gradually become a hot spot in the field of research. In the application of intelligent surveillance, how to effectively identify the posture of certain posture behavior has strong practicability and research value, so it has become one of the key research directions of intelligent video surveillance. Because the monitoring scene is often very complex, in order to obtain a good attitude detection results, we need a set of effective attitude detection scheme. Based on the mature application of Hidden Markov Model in speech recognition, a new attitude detection technique based on Hidden Markov Model is proposed in this paper. The scheme includes the following steps: foreground detection, foreground filtering, target tracking, feature extraction and attitude sequence acquisition. According to the fact that human posture has the characteristics of spatio-temporal continuity, this scheme selects the space-time information in the process of human posture as the feature, and analyzes the human posture. The method of feature analysis based on time and space is used to depict human posture effectively. For the time information of human posture in two frames of image, the method of optical flow is used to extract the motion information of human body attitude, and the amplitude and phase distribution is taken as the motion feature. For the spatial information of a single frame image, that is, the movement of human posture at a certain time, a new method of human posture matching is proposed based on a fast gradient response graph algorithm. The key action of the local region is used to replace the whole human posture, and the complexity of the matching process is simplified. At the same time, in the process of template matching, the hierarchical template matching is established by using different attitude and different angles. This method can improve the accuracy and ensure the detection speed effectively. Finally, based on the temporal and spatial feature information obtained, the state sequence of human posture is obtained based on Hidden Markov Model, which is used as the detection result. In this paper, the technical framework of this scheme is introduced in the form of chapters, and the prototype system of pose detection based on hidden Markov model is implemented by taking photo attitude detection as an example. Experiments show that this method is robust and real-time.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.6

【共引文献】