基于双通道C3D的基建现场人体异常行为识别  

Human Abnormal Behavior Recognition in Infrastructure Site Based on Dual Channel C3D

作　　者：吴冬梅[1] 卢静 蒋瑜 WU Dong-mei;LU Jing;JIANG Yu

机构地区：[1]西安科技大学通信与信息工程学院

出　　处：《信息技术与信息化》2020年第1期28-31,共4页Information Technology and Informatization

摘　　要：异常行为识别在智能监控领域有广泛的应用前景。本文提出一种基于双通道C3D(Convolutional 3D,三维卷积)的行为识别方法,对打架、向下抛物、摔倒、跨越警戒线这四类异常行为以及走路、跑步、工作这三类正常行为进行识别。该方法的一个通道通过提取视频的RGB图像送入C3D网络来获取静态特征;另一个通道通过提取视频的光流图像送入C3D网络来获取动态特征;最后,利用双通道网络在卷积层融合、全连接层融合、混合融合的方法将静态特征与动态特征相结合,对比实验结果表明,最优识别率达到97.7564%,证明了该网络结构在基建现场应用场景中的有效性和可行性。Abnormal behavior recognition has wide application prospects in the field of intelligent monitoring.This paper proposes a behavior recognition method based on dual-channel C3D(Convolutional 3D,three-dimensional convolution),which combats four types of abnormal behaviors:fight,throw,trip,span and three types of normal behaviors:walk,run,and work.One channel of the method obtains the static features by extracting the RGB images of the video and sends them to the C3D network;the other channel obtains the dynamic features by extracting the optical flow images of the video and sends it to the C3D network.Finally,the dual-channel network is used to fuse the convolutional layer,the fully connected layer,and the hybrid fusion method to combine static features with dynamic features.Comparative experimental results show that the optimal recognition rate reaches 97.7564%,which proves the effectiveness and feasibility of the network structure in the field application scenario of infrastructure construction.

关 键 词：异常行为识别 深度学习 C3D卷积神经网络 网络模型融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]