LSCN: Concerning Long and Short Sequence Together for Action Recognition

YANG Ke; WANG Jing-yu; QI Qi; SUN Hai-feng; WANG Jing; LIAO Jian-xin

doi:10.3969/j.issn.0372-2112.2020.03.012

您当前的位置：

首页 >

文章列表页 >

LSCN: Concerning Long and Short Sequence Together for Action Recognition

更新时间：2025-07-16

- LSCN: Concerning Long and Short Sequence Together for Action Recognition
- Acta Electronica Sinica Vol. 48, Issue 3, Pages: 503-509(2020)
- 作者机构：
  
  北京邮电大学网络与交换国家重点实验室,北京,100876
- 作者简介：
- 基金信息：
- DOI：10.3969/j.issn.0372-2112.2020.03.012
  CLC： TP391.4
- Published Online：25 March 2020，
  
  Published：2020
- 稿件说明：
移动端阅览
YANG Ke, WANG Jing-yu, QI Qi, et al. LSCN: Concerning Long and Short Sequence Together for Action Recognition[J]. Acta Electronica Sinica, 2020, 48(3): 503-509.
DOI：

YANG Ke, WANG Jing-yu, QI Qi, et al. LSCN: Concerning Long and Short Sequence Together for Action Recognition[J]. Acta Electronica Sinica, 2020, 48(3): 503-509. DOI： 10.3969/j.issn.0372-2112.2020.03.012.

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.

Abstract

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

Keywords

references

Views

312

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

Variable Horizon Multi-Directional Scanning Method for Time Series Anomaly Detection

Object Detection Based on EIMYOLO for High-Resolution Remote Sensing Images

Multi-Dimensional Dynamic Topology Learning Graph Convolution for Skeleton-Based Action Recognition

Method on Alzheimer’s Disease Classification Utilizing Deep Learning and Hippocampus Heterogeneous Feature Fusion

Fast Small Object Detection Method with Scale-Sensitivity Loss and Feature Fusion

Related Author

HUANG Yu-zhe

GUAN Yong-yuan

WEI Song-jie

CAO Feng

ZENG Ke-wen

LI De-yu

LUO Xi-zhao

TAO Chong-ben

Related Institution

School of Computer Science and Engineering, School of Cyber Science and Engineering, Nanjing University of Science and Technology

School of Information and Technology, Shanxi University

School of Computer Science and Technology, Soochow University

School of Electronics and Information Engineering, Suzhou University of Science and Technology

School of Information Engineering， Jiangxi University of Science and Technology

⁰