面向助老行为识别的三维卷积神经网络设计  

Design of 3D Convolutional Neural Network for Action Recognition for Helping the Aged

作　　者：李秀智 张冉[1,2] 贾松敏 LI Xiuzhi;ZHANG Ran;JIA Songmin(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]数字社区教育部工程研究中心,北京100124

出　　处：《北京工业大学学报》2021年第6期589-597,共9页Journal of Beijing University of Technology

基　　金：北京市教育委员会科技计划资助项目(JZ041001201701)。

摘　　要：针对室内老人跌倒问题,提出一种室内人体跌倒行为识别方法.首先,提出基于卷积核分解与分组卷积的轻量化3D网络;之后融合浅层2D子网络与轻量化3D子网络,并采用随机滑动组合采样策略改进3D卷积行为识别网络.为进一步提高网络泛化性能,对视频帧进行视觉显著性检测,通过加强背景纹理与人物行为之间关联性提高真实场景识别准确度.实验结果表明:该网络参数量为6.9×106,时间复杂度降低至8.04×109;实现算法在室内跌倒行为识别任务上达到81.5%的准确度.To solve the problem of action recognition in indoor environment,a method for human falling recognition in indoor environment was proposed.First,a lightweight 3D network,which uses grouping convolution and factorization to lighten the network structure for action classification,was proposed.Then 2D subnetworks and lightweight 3D sub-networks were fused to improve behavior recognition network based on the 3D convolution.Finally,visual saliency detection was performed on video frames to improve the accuracy of real scene recognition by enhancing the correlation between background texture and human behavior.Results show that the network’s parameter is reduced to 6.9×106 and the floating point of operations is reduced to 6.9×109.The algorithm achieves 81.5%accuracy in the task of indoor fall behavior recognition.

关 键 词：行为识别 跌倒检测 3D卷积神经网络 视觉显著性 卷积核分解 分组卷积 

分 类 号：U461[机械工程—车辆工程] TP308[交通运输工程—载运工具运用工程]