Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (20): 180-187.DOI: 10.3778/j.issn.1002-8331.2006-0154

Previous Articles     Next Articles

Research on Pedestrian Gait Recognition Based on Multi-scale Feature Transfer Learning

XU Jian, HUANG Lei, CHEN Qianqian, LU Zhen, WU Shupei   

  1. School of Electronic Information, Xi’an Polytechnic University, Xi’an 710048, China
  • Online:2021-10-15 Published:2021-10-21

基于多尺度特征迁移学习的步态识别研究

徐健,黄磊,陈倩倩,陆珍,吴曙培   

  1. 西安工程大学 电子信息学院,西安 710048

Abstract:

In order to solve the problem of small sample size of pedestrian gait datasets and the problem of insufficient feature description of single-feature or multiple feature fusion gait recognition algorithm, a pedestrian gait recognition method based on multi-scale feature deep migration learning is proposed. The steps of the algorithm include:Firstly, improving the VGG-16 network by removing the last Maximum pooling Layer(Maxpool Layer) in the network and fusing the Spatial Pyramid Pooling network structure(SPP) to obtain multiscale information of the pedestrian Gait Energy Image(GEI). Then the network model is pretrained by using the Imagenet dataset, and the feature extraction ability is transferred to the pedestrian gait recognition network model. It uses pedestrian gait sample set to fine-tune the network, modifies the full connection layer parameters in the network, and finally applies to the pedestrian gait recognition research. This method achieves 95.7% recognition accuracy on the CASIA-B gait dataset at the CAS Institute of Automation. The method has been demonstrated higher recognition rate than single-feature or multiple feature fusion gait recognition algorithm. It is shown that this method has better recognition performance.

Key words: gait recognition, transfer learning, gait energy image, spatial pyramid pooling, multi-scale features

摘要:

为了解决行人步态数据集样本量较少、单特征或多特征融合的步态识别算法特征描述不足的问题,提出了一种基于多尺度特征深度迁移学习的行人步态识别方法。该算法步骤包括:改进VGG-16网络,去除网络中最后一个最大池化层(Maxpool Layer),融合空间金字塔池化网络结构(SPP)获取行人步态能量图(GEI)的多尺度信息,利用Imagenet数据集预训练此网络模型,将提取特征能力迁移至行人步态识别网络模型中,采用行人步态样本集微调网络,修改网络中的全连接层参数,应用于行人步态识别研究。该方法在中科院自动化研究所的CASIA-B步态数据集上的识别精度达到了95.7%,与单一步态特征的步态识别方法以及融合多种步态特征的识别方法相比,步态识别率有了明显提升,表明该方法有更好的识别性能。

关键词: 步态识别, 迁移学习, 步态能量图, 空间金字塔池化, 多尺度特征