复杂环境下行为识别特征提取方法研究

发布时间：2018-07-10 13:07

本文选题：行为识别 + 特征提取　；参考：《东南大学》2016年博士论文

【摘要】：行为识别是计算机视觉领域的研究热点之一,在虚拟现实、增强现实、智能监控、移动分析和基于内容的视频检索等领域具有广阔的应用前景。现有的简单的、日常标准动作的行为识别技术逐渐成熟,而对于复杂环境下的行为识别,如何提高识别率和效率成为了行为识别的研究焦点。为了提高识别率,目前行为识别的研究主要针对特征提取方法和分类方法两个方面,特征是分类器的输入数据,其性能直接影响分类器的分类效果,进而影响行为识别的识别率和效率。本论文主要研究复杂环境下行为识别的特征提取方法,以提高识别率和效率,主要工作与创新如下：1)提出一种基于移动边界采样的底层特征提取方法：移动边界采样为光流(optical flow)的不连续区域,并将其应用到改进型稠密轨迹上,减少一些不必要的移动轨迹,在保证识别率的前提下提高了效率并且节省了存储空间。仿真结果表明,在应用移动边界采样方法时,识别率与改进型稠密轨迹相当,复杂环境下的轨迹数减少了一半,存储效率提高了一倍。2)提出一种基于线性降维技术的中层特征提取方法：在行为识别的特征提取过程中多次使用线性降维算法,即在FV (fisher vector)提取之前,对底层描述器(extractor)进行一次降维,在FV进入分类器之前,采用线性的降维技术对其进行降维,从而提高效率和识别率。仿真结果表明,采用降维技术比不采用降维技术的测试和训练时间至少可降低2个数量级,并且UCF11识别率提高到92.51%, HMDB51提高到56.47%。3)设计一种快速分类方法和融合策略：快速分类方法用的是极限学习机,融合策略分为前融合策略和后融合策略,前融合策略分为底层描述器(位置1)融合和中层描述器(位置2)融合,后融合策略主要是通过六种不同的规则进行识别率的分析和比较,提高了大类行为识别的效率和识别率。实验结果表明,ELM的训练时间和测试时间至少比SVM降低了10倍,位置2融合和后融合的求积、求和规则对应的是最稳定并且识别率最高的融合。基于以上研究,提出一种基于移动边界采样的改进型稠密轨迹和多次使用降维方法的行为识别架构,进一步提高行为识别的效率和识别率。设计了一种基于非全能量下的主成分分析的降维方法,根据能量比方法、特征值大于平均特征值、特征值大于1、broken_stick和改进型broken_stick方案来选择特征值的个数实现识别率的最大化。实验结果显示,在非全能量下,broken_stick和改进型的broken stick方案获得较好的识别率,比全能量下提高了0.5%,并且效率也得到了提高。
[Abstract]:Behavior recognition is one of the hotspots in the field of computer vision. It has a broad application prospect in the fields of virtual reality, augmented reality, intelligent monitoring, mobile analysis and content-based video retrieval. The existing simple, daily standard action behavior recognition technology is gradually mature, and how to improve the recognition rate and efficiency of behavior recognition in complex environment has become the focus of behavior recognition research. In order to improve the recognition rate, the current research on behavior recognition mainly focuses on two aspects: feature extraction method and classification method. The feature is the input data of the classifier, and its performance directly affects the classification effect of the classifier. Furthermore, it affects the recognition rate and efficiency of behavior recognition. In this paper, we mainly study the feature extraction method of behavior recognition in complex environment to improve the recognition rate and efficiency. The main work and innovations are as follows: (1) A method of feature extraction based on moving boundary sampling is proposed: the moving boundary is sampled as a discontinuous region of optical flow (optical flow), and applied to the improved dense trajectory. It reduces some unnecessary moving trajectories, improves efficiency and saves storage space while ensuring recognition rate. The simulation results show that the recognition rate is equivalent to that of the improved dense trajectory, and the number of tracks in complex environment is reduced by half when the moving boundary sampling method is applied. (2) A middle level feature extraction method based on linear dimensionality reduction is proposed: linear dimensionality reduction algorithm is used many times in the feature extraction process of behavior recognition, that is, before FV (fisher vector) extraction, The dimensionality of the underlying descriptor (extractor) is reduced once, and the linear dimensionality reduction technique is used to improve the efficiency and recognition rate before the FV enters the classifier. The simulation results show that the testing and training time of using dimensionality reduction technique can be reduced by at least two orders of magnitude compared with that without dimensionality reduction technique. And the recognition rate of UCF11 is raised to 92.51 and HMDB51 is raised to 56.47.3) A fast classification method and fusion strategy are designed: the fast classification method uses the extreme learning machine, and the fusion strategy is divided into pre-fusion strategy and post-fusion strategy. The former fusion strategy is divided into the bottom descriptor (position 1) fusion and the middle level descriptor (position 2) fusion. The post fusion strategy mainly analyzes and compares the recognition rate through six different rules, which improves the efficiency and recognition rate of large class behavior recognition. The experimental results show that the training time and test time of ELM are at least 10 times lower than that of SVM. The quadrature of position 2 fusion and post fusion is the most stable fusion with the highest recognition rate. Based on the above research, an improved behavior recognition architecture based on moving boundary sampling and multiple dimensionality reduction methods is proposed to improve the efficiency and recognition rate of behavior recognition. A dimensionality reduction method based on principal component analysis (PCA) with incomplete energy is designed. According to the energy ratio method, the eigenvalue is greater than the average eigenvalue, and the eigenvalue is greater than 1kW / s stick and the improved broken_stick scheme to select the number of eigenvalues to maximize the recognition rate. The experimental results show that the recognition rate of the proposed scheme is better than that of the full energy broken stick scheme, and the efficiency is also improved.
【学位授予单位】：东南大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】