中文核心期刊
CSCD来源期刊
中国科技核心期刊
RCCSE中国核心学术期刊
重庆交通大学学报（自然科学版）

重庆交通大学学报（自然科学版） ›› 2022, Vol. 41 ›› Issue (09): 9-17.DOI: 10.3969/j.issn.1674-0696.2022.09.02

• 交通+大数据人工智能 • 上一篇    下一篇

基于多粒度特征分割的车辆重识别算法

蓝章礼，王超，杨晴晴，金豪   

  1. (重庆交通大学 信息科学与工程学院，重庆 400074)
  • 收稿日期:2021-06-28 修回日期:2021-08-19 发布日期:2022-09-30
  • 作者简介:蓝章礼（1973—），男，重庆人，教授，博士，主要从事图像处理、交通信息化的研究。E-mail: lzl7309@126.com 通信作者：王超（1997—），男，重庆人，硕士研究生，主要从事图像处理与机器视觉的研究。E-mail: w_allen@mails.cqjtu.edu.cn
  • 基金资助:
    重庆市技术创新与应用发展项目（cstc2020jscx-msxmX0073）

Vehicle Re-identification Algorithm Based on Multi-granularity Feature Segmentation

LAN Zhangli, WANG Chao, YANG Qingqing, JIN Hao   

  1. (College of Information Science and Engineering, Chongqing Jiaotong University, Chongqing 400074, China)
  • Received:2021-06-28 Revised:2021-08-19 Published:2022-09-30

PDF

摘要： 针对车辆重识别任务中局部特征提取不充分和潜在显著性局部特征易被掩盖的问题，提出一种基于多粒度特征分割的算法。该算法采用可实现跨通道间信息交互的ResNeSt-50作为骨干网络提取初级特征，并将骨干网络复制成三个独立的分支，对输出的特征图分别沿纵向、横向和通道方向进行多粒度分割以提取到区分性局部特征。为进一步增强网络提取判别性特征信息的能力，又在ResNeSt-50的每个split-attention block中嵌入了空间注意力模块。研究结果表明：算法在VeRi-776数据集上的mAP、Rank-1、Rank-5指标分别达到85.92%、97.67%、98.53%；在VehicleID数据集的三个测试集上，Rank-1指标分别达到了88.36%、84.19%、78.89%，优于现有大部分主流算法，研究结果表明该算法具有先进性和有效性。

关键词: 交通工程；车辆重识别； ResNeSt-50；多粒度特征分割；空间注意力

Abstract: In order to solve the problem that local feature extraction was insufficient and potentially significant local features were easy to be concealed in vehicle re-identification task, an algorithm based on multi-granularity feature segmentation was proposed. In the proposed algorithm, ResNeSt-50 which could realize the information exchange between channels was adopted as the backbone network to extract primary features. And the backbone network was copied into three independent branches, and the output feature map was segmented by multi-granularity along the vertical, horizontal and channel directions to extract the differentiated local features. In order to further enhance the ability of the network to extract discriminant feature information, the spatial attention module was embedded in each split-attention block of ResNeSt-50. The research results show that The mAP, Rank-1 and Rank-5 indexes of the proposed algorithm on the VeRi-776 dataset reach 85.92%, 97.67% and 98.53% respectively; on the three test sets of the VehicleID dataset, the Rank-1 index reaches 88.36%, 84.19% and 78.89% respectively, which is better than that of most of the existing mainstream algorithms. The research results show that the proposed algorithm is advanced and effective.

Key words: traffic engineering; vehicle re-identification; ResNeSt-50; multi-granularity feature segmentation; spatial attention

中图分类号: