基于改进GaitSet的跨视角步态识别方法  

Cross-view Gait Recognition Based on Improved GaitSet

作　　者：孟洪杰 杜延墨 Meng Hongjie;Du Yanmo(The 33rd Research Institute of China Electronics Technology Group Corporation,Taiyuan Shanxi 030032,China;Key Laboratory of Instrumentation Science and Dynamic Testing of Ministry of Education,North University University,Taiyuan Shanxi 030051,China)

机构地区：[1]中国电子科技集团公司第三十三研究所,山西太原030032 [2]中北大学仪器科学与动态测试教育部重点实验室,山西太原030051

出　　处：《机械管理开发》2025年第1期268-270,276,共4页Mechanical Management and Development

摘　　要：针对现有的步态识别模型识别准确率不够高、特征提取层次不足、时序信息提取不充分等问题,提出了一种改进的时空特征融合GaitSet跨视角步态识别方法。该方法利用卷积神经网络从步态序列中提取空间特征,结合多种尺寸的卷积核和膨胀卷积技术来获取多尺度特征。在特征提取阶段引入残差单元以增强深层特征的提取能力。采用长短期记忆网络捕捉时序信息,并在特征融合层将时空特征融合。利用水平金字塔映射进一步提取多种层次的时空特征。在CASIA-B数据集上的实验结果表明,该方法在正常行走、携带包裹和穿着外套三种场景下的全方位平均准确率分别达到95.7%、90.6%和79.2%,相比GaitSet模型分别提高了0.7、3.4和8.8个百分点,验证了方法的有效性。Aiming at the problems that the existing gait recognition models do not have high enough recognition accuracy,insufficient level of feature extraction,and insufficient extraction of temporal information,an improved spatio-temporal feature fusion GaitSet cross-view gait recognition method is proposed.The method utilizes a convolutional neural network to extract spatial features from gait sequences,and combines multiple sizes of convolutional kernels and expansion convolution techniques to obtain multi-scale features.Residual units are introduced in the feature extraction stage to enhance the extraction of deep features.The long and short-term memory network is used to capture the temporal information,and the temporal and spatial features are fused in the feature fusion layer.The horizontal pyramid mapping is utilized to further extract spatio-temporal features at multiple levels.The experimental results on CASIA-B dataset show that the all-around average accuracy of the method reaches 95.7%,90.6%and 79.2%in three scenarios of normal walking,carrying a parcel and wearing a jacket,respectively,which are improved by 0.7,3.4 and 8.8 percentage points compared with the GaitSet model,and the effectiveness of the method is verified.

关 键 词：GaitSet算法 步态识别 残差网络 膨胀卷积 时空特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]