人体动作视频识别TSM优化算法  

TSM optimization algorithm for human motion video recognition

作　　者：马仲甜 李金泉 杨庆江 Ma Zhongtian;Li Jinquan;Yang Qingjiang(School of Electronic&Information Engineering,Heilongjiang University of Science&Technology,Harbin 150022,China)

机构地区：[1]黑龙江科技大学电子与信息工程学院,150022

出　　处：《黑龙江科技大学学报》2024年第5期817-822,共6页Journal of Heilongjiang University of Science And Technology

基　　金：黑龙江省极薄煤层智能开采关键技术攻关与示范项目(2021ZXJ02A02)。

摘　　要：针对人体动作视频识别的结果稳定性差和泛化能力弱等影响因素,提出一种TSM2.0优化算法。采用Resnet50网络作为主干网络进行特征提取,同时引入非局部操作算子来优化TSM网络,提升网络模型整体性能,实现对网络结构的优化。结果表明,TSM2.0算法训练得到的损失函数曲线下降状态更加平稳,训练准确率均达到92%以上,有效减少了损失函数曲线前期出现的过拟合现象,进而提高了模型的稳定性和泛化能力。该研究在进行人体动作视频识别时具有较好的性能。This paper is aimed at addressing the influence factors including poor stability and weak generalization ability on human body action video recognition,and proposes a TSM2.0 optimization algorithm.This is achieved by using Resnet50 network as the backbone network for feature extraction;introducing non-local operators to optimize TSM network,improve the overall performance of the network model,and optimize the network structure.The results show that the downward state of the loss function curve obtained by TSM2.0 algorithm is more stable,and the training accuracy rate is more than 92%,which effectively reduces the overfitting phenomenon of the loss function curve in the early stage,and improves the stability and generalization ability of the model.This research has better performance in human motion video recognition.

关 键 词：视频识别 深度学习 TSM 非局部操作算子 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]