基于改进稠密轨迹的人体行为识别方法研究

发布时间：2018-09-01 13:29

【摘要】：人体行为识别是机器视觉和人工智能领域的研究热点,在视频监控、基于内容的视频检索、人机交互、智能交通等领域有着广泛的应用前景。由于人体运动的非刚性、背景的复杂性、人与人之间的相互遮挡以及相机运动等,人体行为识别成为一个极具挑战性的课题。近年来,研究人员在相关领域取得了很大进展,特别是基于稠密轨迹的人体行为识别方法取得了较之前方法更好的识别效果。本文提出一种基于改进稠密轨迹的人体行为识别方法。首先,在多尺度空间采样稠密光流场提取特征点,通过中值滤波跟踪特征点形成轨迹。现实场景中,视频大都存在相机运动,导致大量的轨迹存在于背景中,给人体行为的识别带来干扰。本文通过显著性检测提取显著性区域,并结合背景轨迹方向一致信息,抑制相机运动的影响。然后,以轨迹为中心建立时空管道,进一步将时空管道分割为时空网格,在时空网格中提取梯度方向直方图、光流直方图和运动边界直方图,联合每类特征形成最终的描述符。为了降低计算复杂度,使用主成分分析法(PCA)对每类特征进行降维。最后,每一类特征用一个Fisher向量表示,并用线性SVM进行分类。通过对稠密轨迹算法进行改进,形成完整的行为识别方法。在公共数据集KTH和YouTube上的实验结果表明,本文方法能够有效地抑制相机运动带来的影响,提高行为识别的准确率。
[Abstract]:Human behavior recognition is a hot topic in the field of machine vision and artificial intelligence. It has a wide application prospect in video surveillance, content-based video retrieval, human-computer interaction, intelligent transportation and so on. Due to the non-rigidity of human motion, complexity of background, mutual occlusion between people and camera motion, human behavior recognition has become a very challenging task. In recent years, researchers have made great progress in related fields, especially the human behavior recognition method based on dense trajectory has achieved better results than previous methods. In this paper, a human behavior recognition method based on improved dense trajectory is proposed. Firstly, feature points are extracted from dense optical flow field in multi-scale space, and feature points are tracked by median filter. In the real scene, most of the videos have camera motion, which leads to a large number of tracks in the background, causing interference to the recognition of human behavior. In this paper, the salience region is extracted by salience detection, and the influence of camera motion is suppressed by combining with the consistent information of background trajectory direction. Then, the space-time pipeline is built with the locus as the center, and the space-time pipeline is further divided into spatio-temporal grids. Gradient direction histogram, optical flow histogram and moving boundary histogram are extracted from the spatio-temporal grid to form the final descriptor by combining each class of features. In order to reduce computational complexity, principal component analysis (PCA) is used to reduce the dimension of each feature. Finally, each class of features is represented by a Fisher vector and classified by linear SVM. By improving the dense trajectory algorithm, a complete behavior recognition method is formed. Experimental results on common data sets KTH and YouTube show that the proposed method can effectively suppress the effect of camera motion and improve the accuracy of behavior recognition.
【学位授予单位】：天津工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】