基于多特征融合卷积的步态识别算法研究  

GAIT RECOGNITION ALGORITHM BASED ON MULTI-FEATUREFUSION CONVOLUTION

作　　者：杨鹏 应娜[1] 李怡菲 Yang Peng;Ying Na;Li Yifei(School of Telecommunication Engineering,Hangzhou Dianzi University,Hangzhou 310018,Zhejiang,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310018

出　　处：《计算机应用与软件》2024年第1期139-145,共7页Computer Applications and Software

基　　金：浙江省自然科学基金项目(LY16F010013)。

摘　　要：针对GaitSet算法中主干网络学习能力和分类能力较弱,提出基于多特征融合卷积网络的步态识别算法(MFFC-GaitSet)。算法通过多特征融合卷积重建GaitSet网络增强网络学习能力,同时对三元组损失函数进行平滑优化;利用形态学处理对步态轮廓图进行修补。算法在Casia-B数据集上进行验证,步态识别精度达到85.811%,提高2.6%;模型权重仅增加6%。算法可以有效减少复杂环境对步态识别的负面影响,实现复杂环境下高精度的步态识别。实验结果表明,方法能够实现较为精确的步态识别,并具有较佳的鲁棒性和泛化能力。Aimed at the weak learning and classification ability of the backbone network in the GaitSet algorithm,the gait recognition algorithm based on the multi-feature fusion convolution(MFFC-GaitSet)is proposed.The algorithm reconstructed the GaitSet network by multi-feature fusion convolution to enhance the network learning ability,and smoothed and optimized the ternary loss function.The gait contour map was repaired by morphological processing.The algorithm was validated on the Casia-B dataset and achieved a gait recognition accuracy of 85.811%,with the increase of 2.6%.The model weight was increased by only 6%.The algorithm could effectively reduce the negative influence of complex environment on gait recognition and achieve high-precision gait recognition in complex environment.The experimental results show that the method can achieve more accurate gait recognition with better robustness and generalization ability.

关 键 词：步态识别 多特征融合 形态学处理 三元组平滑优化 Casia-B数据集 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]