基于多维分割和特征融合的车辆重识别算法  

Vehicle Re-identification Algorithm Based on Multidimensional Segmentation and Feature Fusion

作　　者：张迪[1] 王国栋[1] ZHANG Di;WANG Guodong(College of Computer Science&Technology,Qingdao University,Qingdao 266071,China)

机构地区：[1]青岛大学大学计算机科学技术学院,山东青岛266071

出　　处：《青岛大学学报（工程技术版）》2024年第1期23-30,52,共9页Journal of Qingdao University(Engineering & Technology Edition)

基　　金：青岛市自然科学基金资助项目(RZ2300002511)。

摘　　要：针对车辆重识别算法受类内差异性和类间相似性干扰,细粒度特征提取不充分的问题,本文提出一种融合全局与局部信息,基于卷积和自注意力机制的车辆重识别算法。采用双分支架构对车辆语义进行编码,更好地捕捉车辆信息,为了增强局部语义信息的提取,针对不同维度上的注意力模块,通过设计全局局部交互模块,将全局信息更好地融入局部特征。在公开数据集VeRi 776与VehicleID数据集进行实验验证,实验结果证明所提方法的优越性。To address the challenges of intra-class variability and inter-class similarity interference in vehicle re-identification(Re-ID)algorithms,leading to insufficient extraction of fine-grained features,this paper proposes a vehicle Re-ID algorithm that integrates global and local information based on convolutional and self-attention mechanisms.The algorithm employs a dual-branch architecture for semantic encoding of vehicles,capturing vehicle information more effectively.Additionally,to enhance the extraction of local semantic information,attention modules operating on different dimensions are introduced.Through the design of a global-local interaction module,global information is seamlessly integrated into local features.Finally,experimental validation on the publicly available VeRi-776 and VehicleID datasets demonstrates the superiority of the proposed method.This research exhibits robust advantages in real-world vehicle reidentification scenarios.

关 键 词：多维分割 信息融合 车辆重识别 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]