基于特权信息的人类行为识别研究

发布时间：2019-04-02 17:40

【摘要】：近年来,人类行为识别作为计算机视觉的一个重要领域,广泛应用于人机交互、视频监控、医疗保健和视频检索等场景,为人类的生产和生活带来了便利,具有广阔的应用前景,而鲁棒的识别方法是其大规模应用于实际场景的前提。本文主要提出了两种鲁棒的人类行为识别学习框架,即:基于特权信息的人类行为识别和结合特权信息的域自适应人类行为识别。经典人类行为识别方法通过对训练和测试样本提取一致的视觉特征,学习分类器实现行为分类,并未涉及特权信息。本文提出了一种基于特权信息的人类行为识别框架,针对训练样本提取鲁棒的特权信息,辅助固有视觉特征学习新的分类器。在人体动作识别层面,本文将人体骨架李群深度特征作为特权信息,并在两个公开数据集UTKinect-Action3D和Florence 3D Actions上验证了其有效性;在人类活动识别层面,本文将Web周边文本描述特征作为特权信息,并在Web数据集Flickr上验证了其鲁棒性。经典人类行为识别方法考虑训练和测试样本具有相同的特征空间和共同的数据分布,但在现实应用中获取有标签的训练样本耗时耗力,受限于场景的变换、人物的不同等原因,很难得到满足上述条件的充足样本。本文提出了一种结合特权信息的域自适应人类行为识别方法,允许训练和测试样本来自不同的领域,算法旨在最小化源域和目标域样本之间的分布差异,将此标准结合图正则稀疏编码纳入目标函数中,使得新的稀疏表达对数据分布差异具有鲁棒性,同时在源域中提取特权信息,辅助固有视觉特征学习新的鲁棒分类器。该算法在两组不同域数据集组合FlickrCCV及FlickrKodak的实验上取得了良好的效果。
[Abstract]:In recent years, as an important field of computer vision, human behavior recognition has been widely used in human-computer interaction, video surveillance, health care and video retrieval, which has brought convenience to human production and life. It has a broad application prospect, and robust identification method is the premise of its large-scale application to the actual scene. In this paper, two robust learning frameworks for human behavior recognition are proposed, namely, human behavior recognition based on privileged information and domain adaptive human behavior recognition based on privileged information. Classical human behavior recognition methods extract consistent visual features from training and test samples and learn classifiers to achieve behavior classification without involving privileged information. In this paper, a human behavior recognition framework based on privileged information is proposed, which extracts robust privileged information for training samples and assists in learning new classifiers based on intrinsic visual features. In the aspect of human motion recognition, this paper takes the depth feature of human skeleton Li Qun as privileged information, and validates its effectiveness on two open datasets, UTKinect-Action3D and Florence 3D Actions. At the level of human activity recognition, this paper takes the Web peripheral text description feature as privileged information and validates its robustness on the Web dataset Flickr. The classical human behavior recognition method considers that training and testing samples have the same feature space and common data distribution, but it takes time and energy to obtain labeled training samples in practical applications, which is limited by the transformation of scene and the difference of characters, and so on. It is difficult to obtain sufficient samples to satisfy the above conditions. In this paper, a domain adaptive human behavior recognition method combining privileged information is proposed, which allows training and testing samples to come from different fields. The algorithm aims to minimize the distribution difference between source and target domain samples. This standard is combined with graph regular sparse coding into the objective function, which makes the new sparse expression robust to the difference of data distribution. At the same time, the privilege information is extracted from the source domain to assist the intrinsic visual feature learning of the new robust classifier. The proposed algorithm has achieved good results in the experiment of combining FlickrCCV and FlickrKodak with two sets of data sets from different domains.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】