LSCN:一种用于动作识别的长短时序关注网络  

LSCN:Concerning Long and Short Sequence Together for Action Recognition

作　　者：杨珂 王敬宇[1] 戚琦[1] 孙海峰[1] 王晶[1] 廖建新[1] YANG Ke;WANG Jing-yu;QI Qi;SUN Hai-feng;WANG Jing;LIAO Jian-xin(State Key Laboratory of Networking and Switching Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学网络与交换国家重点实验室,北京100876

出　　处：《电子学报》2020年第3期503-509,共7页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61771068,No.61671079);北京市自然科学基金(No.4182041)。

摘　　要：相较于图像分析,如何分析时序信息是动作识别中的一个主要问题.大多数先前的方法,如3D卷积网络、双流卷积网络,仅使用包含全局时域信息的特征作为视频的表征,忽略了局部时序特征的重要性.考虑到这样的问题,本文提出一种基于时序交互感知模块的长短时序关注网络——Long and Short Sequence Concerned Networks(LSCN),融合不同的时序信息,利用不同卷积层时序特征的交互加强对不同时序长度的动作实例的表示,兼顾长短动作实例对时序信息的需求.实验结果表明,基于3D ResNext101的LSCN在两个公共数据集(UCF101和HMDB51)上,相较于基础的网络分别有0.4%和2.9%的准确率提升.Compared with image analysis,how to analyze temporal information is a challenging problem in action recognition.Most of the previous methods,such as 3D CNNs(convolutional neural networks)and two-streams CNNs,only used features containing global temporal information as video representation,ignoring the importance of local temporal features.To solve this problem,we propose long and short sequence concerned networks(LSCN)based on temporal interaction perception module,which can combine different temporal information.LSCN makes use of the interactions of temporal features from different convolution layers to enhance the representation of videos and takes into account the needs of temporal information for long and short sequence actions.The results of experiments show that LSCN based on 3D ResNext101 can be generalized in two public datasets(UCF101 and HMDB51).Moreover,compared with the basic network,there are 0.4%and 2.9%accuracy improvements respectively.

关 键 词：动作识别 时序特征 特征融合 人机交互 视频分析 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]