基于三维卷积神经网络的工作流识别方法  

Workflow recognition method based on 3D convolutional neural networks

作　　者：胡海洋[1,2] 丁佳民 胡华[1,2] 陈洁 李忠金[1,2] HU Haiyang 1,2 , DING Jiamin 1,2 , HU Hua 1,2 ,CHEN Jie 1,2 , LI Zhongjin 1,2(1. College of Computer, Hangzhou Dianzi University, Hangzhou 310018, China;2. Key Laboratory of Complex System Modeling and Simulation, Ministry of Education, Hangzhou Dianzi University, Hangzhou 310018, Chin)

机构地区：[1]杭州电子科技大学计算机学院,浙江杭州310018 [2]杭州电子科技大学复杂系统建模与仿真教育部重点实验室,浙江杭州310018

出　　处：《计算机集成制造系统》2018年第7期1747-1757,共11页Computer Integrated Manufacturing Systems

基　　金：国家自然科学基金资助项目(61572162;61272188;61702144);浙江省重点研发计划资助项目(2018C01012);浙江省自然科学基金资助项目(LQ17F020003)~~

摘　　要：鉴于传统的依赖于目标物体检测与跟踪的动作识别方法很难适用于复杂的生产制造环境,为了实现有效的工作流识别,从运动物体的检测与分割、视频序列中多视图特征向量的提取及工人生产动作的分类识别3方面入手,提出基于3D卷积神经网络的工作流识别框架。给出计算模型与相应的算法,并进行了系统的对比实验。通过实验发现,该方法比传统的隐Markov方法和其他方法在识别速度上提升了32%,在识别率上也提升了9%。Owing to the problem that traditional method of action recognition based on object detection and tracking might not be applicable to complex manufacturing environments, to realize workflow recognition effectively, by making research on the moving objects detection and tracking, extracting feature vector from video sequence and classification of actions, a multi-view feature extraction framework based on moving object segmentation based on 3D Convolutional Neural Networks was proposed. The calculation model and the corresponding algorithm along with the systematic comparative experiments were also given. Experiments showed that the proposed method could improve the speed of recognition about 32% and the recognition accuracy about 9% compared with traditional Hidden Markov method and other methods.

关 键 词：智能制造 工作流 行为识别 帧间差分 3维卷积神经网络 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]