Abstract: The traditional human behavior management approaches highly rely on video surveillance,its time costing and mistakable nature makes it a chore in complex manufacturing.To achieve more effective management of human activity recognition and intelligent monitoring in the workshops,a working activity recognition approach was proposed based on human skeleton data.The 3D depth vision sensor was utilized to collect the body skeleton joint position data,and the joint data was normalized using a standardized reconstruction approach.RGB images of spatial-temporal features of working activities were represented.On this basis,a deep convolution neural network model was trained for working activity recognition in both of the spatial and the temporal domain.The experiments on the open MSR-Action 3D data set and our own NJUST3Ddata set using CUDA GPU accelerated computing demonstrated that the proposed approach had high accuracy and practical value.

Key words: depth vision, human activity recognition, skeleton, deep learning, deep convolution neural network model

摘要： 传统的依赖视频监控的人员行为管理方式费时且易产生疏漏,难以适用复杂的生产制造环境,为了实现更加有效的人员行为管理,针对生产车间工作人员行为识别与智能监控问题,提出一种基于人体骨架信息的生产行为识别方法。基于三维深度视觉传感器采集人体骨架关节位置数据,用标准化重构方法对骨架关节数据进行归一化处理,合成人体行为的时空特征RGB图像。在此基础上构建深度卷积神经网络模型,进行时空域的生产行为识别。最后通过CUDA GPU加速环境下面向MSR-Action3D数据集和自建验证数据集NJUST3D进行实验验证,说明所提方法具有较高的准确率和实用价值。

关键词: 深度视觉, 行为识别, 骨架, 深度学习, 深度卷积神经网络模型