基于时空信息的行为识别方法研究
首发时间:2022-05-23
摘要:视频动作分类,又名行为识别、动作识别,是视频理解领域一个基础的研究任务。其目的是对给定的剪辑好的视频片段进行视频动作的分类,与图像领域的图像识别任务类似。作为视频理解方向的基础任务,行为识别任务已经拥有比较长久的研究历史,包括早期基于传统方法的研究以及随着深度学习发展后基于深度学习模型的研究。基于深度学习的研究工作有很多,较为经典的方法有:双流网络和三维卷积神经网络。其中,双流网络的时间分支需要光流图作为输入,而传统光流的提取需要在输入模型前独立进行并且相当耗时。本文提出使用卷积神经网络来自动从视频图像中提取光流信息,并且将提取的光流图输入光流特征提取网络。同时,为了从光流中提取有效的动作信息特征,提出将光流导向特征子网络嵌入光流特征提取网络以获取有效的动作信息特征。改进的方法在公开数据集UCF-101和HMDB-51上取得良好的效率和准确率。
For information in English, please click here
Research on Action Recognition Based on Spatio-Temporal Information
Abstract:Video action classification, also known as behavior recognition and action recognition, is a basic research task in the field of video understanding.Its purpose is to classify the video actions of a given clip, which is similar to the task of image recognition in the image field.As the basic task of video understanding, action recognition task has a long history of research, including early research based on traditional methods and research based on deep learning model with the development of deep learning.There are a lot of research work based on deep learning, the classical methods are: two stream network and three-dimensional convolutional neural network.Among them, the time branch of two stream network needs the optical flow as input, while the extraction of traditional optical flow needs to be carried out independently before the input model and is quite time-consuming.In this paper, a convolutional neural network is proposed to automatically extract optical flow information from video images and input the extracted optical flow image into optical flow feature extraction network. At the same time, in order to extract effective motion information features from optical flow, the optical flow guide feature subnetwork is embedded into the optical flow feature extraction network to obtain effective motion information features. The improved method achieves good efficiency and accuracy on the open dataset UCF-101 and HDB-51.
Keywords: Action Recognition Convolution neural network Two stream network
基金:
引用
No.****
同行评议
勘误表
基于时空信息的行为识别方法研究
评论
全部评论0/1000