人体动作识别中自步学习算法的研究

发布时间：2018-03-17 06:24

本文选题：自步学习　切入点：多视角　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：现如今,关于视频中人体动作识别的应用在人们日常生活中随处可见,比如智能监控、视频检索、虚拟现实等等,因此基于视频的人体动作识别具有非常广阔的发展前景和极高的研究价值。本文围绕这一课题主要提出了三种识别算法,即:基于多视角自步学习的人体动作识别算法、基于多视角课程序列的多任务人体动作识别算法和基于多视角自步学习序列的多任务人体动作识别算法。首先本文对较为新颖独特的自步学习算法进行了深入的研究,提出了一种基于多视角自步学习的人体动作识别算法。该算法除了保留自步学习本身的优良特性外,还考虑了不同视角对自步学习课程序列的影响。通过使用线性规划增强算法将多个视角下的自步学习分类器进行融合,学习出更加适合解决多类复杂人体动作识别问题的综合课程,进而提高了识别效果。其次本文首次将基于课程的多任务学习用于人体动作识别,该算法将课程学习的思想融入多任务学习中,针对任务的难易程度与相关性设计课程序列以学习多任务。在此基础上,本文提出一种基于多视角课程序列的多任务学习算法,从多种视角学习多任务的课程,利用线性规划增强算法将多个视角下的多任务课程模型进行融合,得到一个更加完备的课程序列多任务模型。最后,本文提出一种基于多视角自步学习序列的多任务学习算法。该算法不仅考虑了多个任务的课程设计,同时利用自步学习算法将每个任务内训练样本的课程嵌入到模型学习中,通过设计不同视角下多任务的课程及训练样本的课程,并将多种视角下的课程模型进行融合,最终设计出了一个更加全面的综合课程模型。利用本文提出的三种算法在公开的人体动作数据集上进行对比实验,实验结果表明本文算法具有较高的准确率和鲁棒性。
[Abstract]:Nowadays, the application of human motion recognition in video is widely used in people's daily life, such as intelligent surveillance, video retrieval, virtual reality and so on. Therefore, human motion recognition based on video has a very broad development prospect and high research value. This paper mainly proposes three recognition algorithms around this topic, namely, human motion recognition algorithm based on multi-view self-learning. Multi-task human motion recognition algorithm based on multi-view course sequence and multi-task human motion recognition algorithm based on multi-view self-stepping learning sequence. In this paper, a human body motion recognition algorithm based on multi-view self-learning is proposed. The effects of different perspectives on the sequence of self-learning courses are also considered. By using linear programming enhancement algorithm, the self-learning classifier with multiple angles of view is fused. Learning a comprehensive course that is more suitable for solving the problem of multi-class complex human motion recognition, and then improving the recognition effect. Secondly, this paper applies the curriculum based multi-task learning to human motion recognition for the first time. The algorithm integrates the idea of course learning into multi-task learning, and designs the course sequence to learn multi-task according to the difficulty and correlation of tasks. On this basis, this paper proposes a multi-task learning algorithm based on multi-perspective course sequence. Learning multi-task courses from multiple perspectives, using linear programming enhancement algorithm to fuse the multi-task curriculum model under multiple perspectives to obtain a more complete curriculum sequence multi-task model. Finally, In this paper, a multi-task learning algorithm based on multi-view self-step learning sequence is proposed. This algorithm not only considers the curriculum design of multiple tasks, but also embeds the training samples of each task into the model learning by using the self-learning algorithm. By designing multi-task curriculum and training sample curriculum under different visual angle, and combining the curriculum model of multi-perspective, Finally, a more comprehensive integrated curriculum model is designed and compared with the three algorithms proposed in this paper on the open human action data set. The experimental results show that the proposed algorithm has high accuracy and robustness.
【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;TP181

【相似文献】