基于3D卷积的人体行为识别技术综述

Review of Human Action Recognition Technology Based on 3D Convolution

随着经济社会的发展,视频分析任务越来越受到重视。同时,人体行为识别技术已广泛应用于虚拟现实、视频监控、视频检索等领域。传统的人类动作识别方法使用2D卷积处理输入视频,但2D卷积只能提取空间特征,而基于手工提取的方法在复杂环境下又难以处理。因此,在深度学习和图像分类任务取得成功的大背景下,基于深度学习的双流网络以及可以同时提取时空特征的3D卷积应运而生。3D卷积在最近几年迅速发展,衍生出多种经典架构且每种框架拥有不同的特性,各种框架皆存在各自的优化方法以及提高速度和精度的效果。在总结几种主流3D卷积框架的基础上将其在相应数据集上进行对比分析,可以得到每种框架的优势及弊端,以此扬长避短,寻找与实际情景相适应的最优框架。

With the development of economy and society,tasks of video analysis are getting more and more attention.Meanwhile,human action recognition technology has been widely used in virtual reality,video surveillance,video retrieval,etc.Traditional human action recognition method is to use 2D convolution to process the input video,but 2D convolution can only extract the spatial features.However,the recognition based on manual extraction in complex environments is difficult to handle.Therefore,in the context of the success of deep learning and image classification tasks,a dual-flow network based on deep learning and a 3D convolution that can simultaneously extract temporal and spatial features emerges.3D convolution has developed rapidly in recent years,and has derived a variety of classic architectures,each with different characteristics.Each framework has its own optimization method and the effect of improving speed and accuracy.Based on the summary of several mainstream 3D convolutional frameworks and putting them into corresponding data sets for comparison and analysis,the advantages and disadvantages of each framework can be obtained accordingly,so as to find the optimal framework that is suitable for the actual situation.