电子学报 ›› 2019, Vol. 47 ›› Issue (8): 1661-1668.DOI: 10.3969/j.issn.0372-2112.2019.08.008

• 学术论文 • 上一篇    下一篇

基于深度学习和智能规划的行为识别

郑兴华1, 孙喜庆1, 吕嘉欣1, 鲜征征2, 李磊1   

  1. 1. 中山大学数据科学与计算机学院, 广东广州 510006;
    2. 广东金融学院互联网金融与信息工程学院, 广东广州 510521
  • 收稿日期:2018-05-18 修回日期:2019-01-02 出版日期:2019-08-25 发布日期:2019-08-25
  • 通讯作者: 鲜征征
  • 作者简介:郑兴华 男,1983年生于广东广州.中山大学数据科学与计算机学院博士.主要研究方向为智能规划、数据挖掘、模式识别、神经网络等.E-mail:zhengxh5@mail3.sysu.edu.cn
  • 基金资助:
    广东省自然科学基金(No.2017A030313391);广东省科技厅国际合作项目(No.2017A050501042)

Action Recognition Based on Deep Learning and Artificial Intelligence Planning

ZHENG Xing-hua1, SUN Xi-qing1, LU Jia-xin1, XIAN Zheng-zheng2, LI Lei1   

  1. 1. School of Data and Computer Science, Sun Yat-sen University, Guangzhou, Guangdong 510006, China;
    2. School of Internet Finance and Information Engineering, Guangdong University of Finance, Guangzhou, Guangdong 510521, China
  • Received:2018-05-18 Revised:2019-01-02 Online:2019-08-25 Published:2019-08-25

摘要: 现有行为识别方法在未能持续覆盖造成视频监控盲区所引起行为数据缺失的情况,难以有效实施特征分析、行为分类补全,无法准确识别出智能体完整的行为动作序列.为此,本文提出一种基于深度学习和智能规划的行为识别方法.首先,利用深度残差网络对图像进行分类训练,然后使用递归神经网络对图像特征进行提取深度信息以增强分类效果;其次,运用智能规划的STRIPS (Stanford Research Institute Problem Solver)模型,将深度学习提取的图像特征命题信息转化为规划领域的模型描述文档,并使用前向状态空间搜索规划器推导出完整的行为动作序列.在HMDB51等行为识别公共数据集中,本方法与生成式对抗网络、深度卷积逆向图网络、深度信念网络、支持向量机等同类先进方法相比展现出更好的性能.

关键词: 行为识别, 深度学习, 智能规划, 深度残差网络, 递归神经网络, STRIPS规划模型, 前向状态空间搜索规划器

Abstract: Currently,action recognition methods can hardly carry out feature analysis,behavior classification,and action completion,and are incapable of accurately identifying the complete behavioral action sequence of intelligent agent for the discontinuous and incomplete motion capture,behavioral data missing or even broken in the time dimension,which are resulted from sensor device not being continuous coverage caused by the monitoring blind area.In this regard,we put forward a method of action recognition based on deep learning and artificial intelligence planning.Firstly,a deep learning network is constructed,by which the image is classified and trained using DRN(Deep Residual Network).After that,the extraction depth information of image frame feature for recurrent neural network is trained to enhance the classification effect.Secondly,the STRIPS(Stanford Research Institute Problem Solver)planning model is used to extract the image feature of deep learning,transforming into the description document for domain model,which facilitates deriving the optimal planning solution by means of forward state-space search planner.In the experiment,we exhibit that our method outperforms baselines in the public datasets,e.g.,DCIGN(Deep Convolutional Inverse Graphics Networks),GAN(Generative Adversarial Networks),DBN(Deep Belief Networks),and SVM(Support Vector Machine).

Key words: action recognition, deep learning, artificial intelligence planning, deep residual network, recurrent neural network, STRIPS planning model, forward state-space search planner

中图分类号: