摘要
时序动作检测作为视频理解中的一项基本任务,被广泛应用于人机交互、视频监控、智能安防等领域。基于卷积神经网络,提出了一种改进的编码-解码时序动作检测算法。改进后的算法分两阶段进行:首先,替换特征提取网络,用残差结构网络提取视频帧的深度特征;之后,构建编码-解码时序卷积网络。采用联接的方式进行特征融合,改进上采样的形式,并运用新的激活函数LReLU进行训练,提高网络的检测精度。实验结果表明,所提算法在时序动作检测数据集MERL Shopping和GTEA上取得了优良的效果。
Temporal action detection is a fundamental task in video understanding that is commonly used in the fields of human-computer interaction,video surveillance,intelligent security,and other fields.An improved encoderdecoder temporal action detection algorithm based on the convolutional neural network is proposed.The improved algorithm is applied in two stages:first,the feature extraction network is replaced and the residual structure network is used to extract the deep features of the video frame;and second,the encoder-decoder temporal convolutional network is constructed.The feature fusion is conducted via contact,and the method of upsampling is improved.To improve the detection accuracy of the network,the proposed algorithm employs the appropriate activation function LReLU for training.The experimental results show that the accuracy of the proposed algorithm on the temporal action detection datasets MERL Shopping and GTEA has improved.
作者
王玥
苏寒松
刘高华
Wang Yue;Su Hansong;Liu Gaohua(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2021年第20期455-461,共7页
Laser & Optoelectronics Progress
关键词
光计算
图像处理
动作检测
时序卷积神经网络
深度学习
optics in computing
image processing
action detection
temporal convolutional neural network
deep learning
