基于相对位置编码转换器模块的深度步态识别网络  

Deep Gait Recognition Network Based on Relative Position Encoding Transformer

作　　者：任禹衡 赵云峰 吴闯 REN Yuheng;ZHAO Yunfeng;WU Chuang(Watrix.AI,Beijing 100083,China;North Pipeline Co.,Ltd.,National Pipe Network Group,Beijing 100084,China;The General Staff of Hunan Provinical Corps of the Chinese People’s Armed Police Force,Changsha 410000,China)

机构地区：[1]银河水滴科技(北京)有限公司,北京100083 [2]国家管道网集团北方管道有限责任公司,北京100084 [3]武警湖南总队参谋部,长沙410000

出　　处：《计算机科学》2024年第S02期521-526,共6页Computer Science

摘　　要：步态识别是一种快速发展的远距离生物特征识别技术,在远距离、跨视角和跨着装等多种场景中具有广泛应用和优势。传统的生物特征识别技术,如指纹识别、面部识别等,往往需要近距离或在特定条件下才能有效进行,而步态识别技术则突破了这些限制,使得在更为广泛的环境下进行个体识别成为可能。以往的研究大多采用轻量级的神经网络提取步态特征,并在目前流行的跨视角和跨着装数据集上(如CASIA-B)取得了巨大的进步。然而,实验结果表明,在CASIA-B数据集上简单叠加神经网络的层数将导致识别准确率大幅度下降。基于相对位置编码转换器模块提出了一个深度步态识别网络,旨在避免陷入“局部特征关联”的陷阱,同时使网络能够持续不断地学习步态序列的时序特征。与目前主流的方法相比,所提方法在室内场景(CASIA-B,OUMVLP)和室外场景(Gait3D)步态数据集上都达到了更优的识别准确率,特别在换装任务(CL)上超出基准方法1.9%,实现了85.5%识别准确率。Gait recognition is a rapidly evolving long-range biometric identification technique that has wide applications and advantages in various scenarios,including long distances,non-intrusive setups,and cross-view angles.Traditional biometrics identification technique,such as fingerprint recognition and facial recognition,often require close proximity or specific conditions to be effective,while gait recognition technology breaks through these limitations,making it possible to identify individuals in a wider range of environments.Previous research predominantly employed lightweight neural networks for gait feature extraction and achieved significant progress on popular datasets like CASIA-B,which feature cross-view angles and varying attires.However,experimental results indicate a substantial decline in recognition accuracy when simply stacking neural network layers on the CASIA-B dataset.A deep gait recognition network has been proposed,incorporating the relative position encoding transformer module.This module aims to avoid the pitfall of“local feature association"”and enables continuous learning of temporal features within gait sequences.Compared to current mainstream approaches,the proposed method has garnered enhanced identification precision across indoor environments,as exemplified by the CASIA-B and OUMVLP datasets,alongside outdoor settings typified by the Gait3D dataset.Especially in the task of clothes changing,wherein our method surpasses benchmark approaches by of 1.9%,achieving a recognition rate of 85.5%.

关 键 词：步态识别 自注意力机制 相对位置建模 模式识别 深层网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]