基于深度学习的人体行为识别算法综述

人体行为识别和深度学习理论是智能视频分析领域的研究热点, 近年来得到了学术界及工程界的广泛重视, 是智能视频分析与理解、视频监控、人机交互等诸多领域的理论基础. 近年来, 被广泛关注的深度学习算法已经被成功运用于语音识别、图形识别等各个领域.深度学习理论在静态图像特征提取上取得了卓著成就, 并逐步推广至具有时间序列的视频行为识别研究中. 本文在回顾了基于时空兴趣点等传统行为识别方法的基础上, 对近年来提出的基于不同深度学习框架的人体行为识别新进展进行了逐一介绍和总结分析; 包括卷积神经网络(Convolution neural network, CNN)、独立子空间分析(Independent subspace analysis, ISA)、限制玻尔兹曼机(Restricted Boltzmann machine, RBM)以及递归神经网络(Recurrent neural network, RNN)及其在行为识别中的模型建立, 对模型性能、成果进展及各类方法的优缺点进行了分析和总结.

A Review of Human Action Recognition Based on Deep Learning

Human action recognition is an active research topic in intelligent video analysis and is gaining extensive attention in academic and engineering communities. This technology is an important basis of intelligent video analysis, video tagging, human computer interaction and many other fields. The deep learning theory has been made remarkable achievements on still image feature extraction and gradually extends to the time sequences of human action videos. This paper reviews the traditional design of action recognition methods, such as spatial-temporal interest point, introduces and analyzes different human action recognition framework based on deep learning, including convolution neural network (CNN), independent subspace analysis (ISA) model, restricted Boltzmann machine (RBM), and recurrent neural network (RNN). Finally, this paper summarizes the advantages and disadvantages of these methods.