基于深度学习的动作识别方法研究

发布时间：2019-01-20 12:28

【摘要】：随着计算机软硬件以及互联网的飞速发展,计算机视觉在各个领域得到了广泛的关注。其中,视频中的人体动作识别技术在安防和监控等领域具有重要作用。在视频人体动作识别研究领域,RGB-D摄像机的出现提供了视频的深度信息,同时也为如何处理这些信息带来了挑战。如何挖掘视频中的有效特征一直是动作识别中的难点之一。另一方面,随着深度学习热潮的兴起,计算机视觉中的多项任务不断取得突破性的进展。因此,基于深度学习框架,本文重点在以下几个方面展开研究:首先,本文提出了一种基于深度神经网络的多模态动作识别方法,根据不同模态信息的特性分别采用不同的深度神经网络,并将多种深度网络结合起来,以挖掘视频中人体动作的多模态时空深度特征。其次,本文针对视频场景复杂造成人体动作显著性不高的问题,提出了一种基于视觉显著性的动作识别方法。该方法结合了三维神经网络以及递归神经网络的特点,设计了一种端到端的深度学习网络模型,同时保留视频的空间和时序特征进行动作识别。最后,本文设计并实现了一个基于Web端的动作识别演示系统,用户通过选择待识别视频,经由服务器处理分析并识别,将最终的识别结果分析呈现在用户端,方便用户分析算法的实际运用效果。
[Abstract]:With the rapid development of computer hardware and software and Internet, computer vision has been widely concerned in various fields. Human motion recognition plays an important role in the field of security and surveillance. In the field of video human motion recognition, the appearance of RGB-D camera not only provides the depth information of video, but also brings challenges to how to deal with this information. How to mine effective features in video is always one of the difficulties in motion recognition. On the other hand, with the rise of deep learning, many tasks in computer vision have made great progress. Therefore, based on the depth learning framework, this paper focuses on the following aspects: first, this paper proposes a multi-modal motion recognition method based on depth neural network. According to the characteristics of different modes of information, different depth neural networks are adopted, and a variety of depth networks are combined to mine the multi-modal space-time depth characteristics of human actions in video. Secondly, a motion recognition method based on visual salience is proposed to solve the problem that the human motion is not significant due to the complexity of video scene. This method combines the characteristics of 3D neural network and recurrent neural network, and designs an end-to-end deep learning network model, while preserving the spatial and temporal features of the video for motion recognition. Finally, this paper designs and implements a motion recognition demonstration system based on Web. By selecting the video to be identified, the user analyzes and recognizes the final recognition result through the server, and presents the final recognition result to the user. The practical application effect of the convenient user analysis algorithm.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】