摘要
目前大多数基于双流卷积网络的行为识别方法采用同样的时空网络结构,双流合并时会产生大量的冗余信息,从而降低识别的精确度。对此提出一种基于双流网络的时空异构网络结构。该网络采用两种不同的时空网络结构对行为进行分类。此外,对视频序列的长时间结构采用分段形式进行建模,使整个行为视频的学习变得高效。在UCF101和HMDB51数据集上进行实验,结果证明该时空异构双流网络优于时空同构双流网络。
Most current action recognition methods based on the two-stream convolutional network use the same structure for spatio-temporal networks,and a large amount of redundant information is generated when two streams are merged,thereby reducing the accuracy of recognition.Based on the above problems,this paper proposes a spatiotemporal heterogeneous network structure based on a two-stream network.This two-stream network used two different spatiotemporal network structures to classify actions.In addition,the long-term structure of the video sequence was modeled in segments to make the learning of the entire action video efficient.The experiments were performed on UCF101 and HMDB51 datasets.The results prove that the proposed spatiotemporal heterogeneous two-stream network is superior to the spatiotemporal homogeneous two-stream network.
作者
丁雪琴
朱轶昇
朱浩华
刘光灿
Ding Xueqin;Zhu Yisheng;Zhu Haohua;Liu Guangcan(College of Automation,Nanjing University of Information Science&Technology,Nanjing 210000,Jiangsu,China)
出处
《计算机应用与软件》
北大核心
2022年第3期154-158,共5页
Computer Applications and Software
基金
国家自然科学基金优秀青年基金项目(61622305)
国家自然科学基金青年基金项目(61502238)
江苏省自然科学基金杰出青年基金项目(BK20160040)。