基于轨迹时空匹配的动作识别算法研究

发布时间：2018-01-06 00:06

本文关键词：基于轨迹时空匹配的动作识别算法研究　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：动作识别在诸多领域有广泛的潜在应用,例如智能视频监控,视频检索,人机交互等,近年来已经成为计算机视觉领域的一个研究热点。尽管许多学者在这方面已经做了非常多的工作,然而它仍然是计算机视觉的一个难点。首先,存在因运动速度的不同,视角变化和背景噪声引起的巨大的类内变化。其次,类别的定义建立在动作的概念和语义之上,底层视频特征和高级语义之间存在的偏差使得分类变得非常困难。在动作识别领域已经出现非常多优秀的方法,其中密集轨迹作为一种高鲁棒性的底层特征和Fisher Vector作为一个通用的组合被广泛应用于动作识别,并在许多公共数据集上取得了优异的表现。它构建基于高斯混合模型(Gaussian Mixture Model,GMM)的字典来表示视频中轨迹的分布,然后将其用于对给定视频的轨迹进行编码。由于对视频中的所有轨迹进行无序编码,从而忽略了不同轨迹之间的时空位置关系。为了弥补时空信息的丢失,本文提出了两种不同轨迹匹配算法来捕获丢失的时空位置关系,即:基于KNN的特征匹配策略和基于堆积的特征匹配策略。首先,我们定义了一个轨迹距离来衡量两条轨迹之间的关系。然后,执行本文提出的特征匹配策略对轨迹进行匹配,对于匹配之后的成对轨迹,我们采用平均池策略(average pooling strategy)来进行特征融合得到本文提出的时空配对轨迹(Space-time Pairwise Trajectories,SPT)。然后用 GMM 和 Fisher Vector 对 SPT进行编码。SPT包含了每条轨迹周围的局部时空信息,从而使得SPT比原始的密集轨迹的特征更具有判别性。我们在四个被广泛应用的公共数据集(Olympic Sports,HMDB51,UCF50,UCF101)上进行实验,实验结果证明我们提出的SPT的表现优于当前优秀的方法。
[Abstract]:Action recognition has potential application in a wide variety of fields, such as intelligent video surveillance, video retrieval, human-computer interaction, in recent years has become a hot research topic in the field of computer vision. Although many scholars have done a lot of work in this area, but it is still a difficult problem in computer vision. First of all, due to different velocities, caused by the change of angle change and background noise of the huge class. Secondly, the definitions of the categories based on concepts and semantic actions, partial job exists between the bottom of the video features and the high-level semantic classification have become very difficult. In the field of action recognition has been a very good number of methods, including dense trajectories as low-level features and Fisher Vector a high robustness as a generic combination is widely used in action recognition, and in many public data sets Excellent performance. It builds the Gauss mixture model based on (Gaussian Mixture Model, GMM) the dictionary said in the video distribution path, which is then used for a given video encoding. Because of the track of the disorder on all track in video encoding, which ignores the relationship between different space-time position trajectory for the missing. For temporal information, this paper presents two different path matching algorithm to capture the spatial position relation, namely: loss of KNN feature matching strategy and matching based on the feature of accumulation based on strategy. First, we define a path to measure the distance between two paths. Then, implementation of the proposed feature matching strategy of trajectory matching for paired trajectory matching after, we use the average pool strategy (average pooling strategy) for feature fusion of the time The empty pair (Space-time Pairwise Trajectories, SPT locus). Then GMM and Fisher Vector of SPT encoding.SPT contains local spatio-temporal information around each trajectory, so that the SPT is more discriminative than characteristic trajectory. The original intensive public data in four widely used set (Olympic Sports, HMDB51, UCF50, UCF101) on the experiment, the experimental results show that our proposed method performs better than the current SPT good.

【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】