多尺度步态特征融合识别算法  

Multi-scale feature fusion for gait recognition

作　　者：王遥枝 于雅楠[1] WANG Yaozhi;YU Yanan(School of Information Technology Engineering,Tianjin University of Technology and Education,Tianjin 300222,China)

机构地区：[1]天津职业技术师范大学信息技术工程学院,天津300222

出　　处：《天津职业技术师范大学学报》2025年第1期47-52,共6页Journal of Tianjin University of Technology and Education

基　　金：天津市教委科研计划项目(2021KJ009).

摘　　要：针对Gaitset模型对于不同步态视角的数据在轮廓特征提取和分类能力方面较弱的问题,提出基于卷积网络的多尺度特征融合算法。该算法在原模型的基础上加入Inception模块并对模型中不同深浅特征进行多尺度融合,增加模型深度,提高模型提取特征的能力。此外,为了提高数据在轮廓细节处的表达能力,将尺寸为64×64的原始图像上采样至128×128。在公开数据集CASIA-B上进行实验,结果表明:该方法对普通行走、背包行走和穿大衣行走的识别准确率分别达到了95.6%、91.4%和75.4%,准确率分别提高了0.6%、4.2%、5.0%;与原算法相比,该方法在轮廓提取与步态识别方面具有更好的鲁棒性和泛化性。To address the limitations of the Gaitset model in contour feature extraction and classification for data from different gait perspectives,a multi-scale feature fusion algorithm based on convolutional networks is proposed.The Inception module is integrated into the original model and multi-scale fusion of features at different depths is performed to enhance the model’s depth and its ability to extract features.Additionally,to amplify the representation of contour details in the data,the original image size of 64×64 data is upsampled to 128×128.Experiments conducted on the publicly available dataset CASIA-B demonstrate that this method achieves recognition rates of 95.6%,91.4%,and 75.4%for normal walking,walking with a backpack,and walking while wearing a coat,respectively.The accuracy rates have been improved by 0.6%,4.2%,and 5%,respectively.Compared to the original algorithm,this method exhibits improved robustness and generalization in contour extraction and gait recognition.

关 键 词：步态识别 卷积神经网络 多尺度特征融合 CASIA-B数据集 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]