基于特征鲁棒性增强的多摄像头下车辆识别方法  

Multi-camera Vehicle Recognition Method Based on Feature Robustness Enhancement

作　　者：罗慧诚 汪淑娟[1] Luo Huicheng;Wang Shujuan(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《系统仿真学报》2023年第5期1059-1074,共16页Journal of System Simulation

基　　金：国家自然科学基金(61962032);云南省优秀青年基金(202001AW070003)。

摘　　要：由于多摄像头下的视角变化、复杂环境和姿态差异等因素,同一辆车在不同场景下的图像表现出巨大的外观歧义性,为车辆的身份匹配带来了挑战。为解决这一难题,在transformer框架下提出了一种面向车辆识别的特征鲁棒性增强的方法。基于车辆结构信息是多个摄像头下域不变的信息,设计了一种轮廓特征引导增强结构信息的模块,并提出了一种结构特征感知损失来促进模型对结构信息的融合;通过将车辆的属性信息作为矢量嵌入到transformer框架中,进一步缓解多视角下车辆姿态改变带来的影响。实验结果表明,所提方法在数据集VeRi-776和VehicleID上均展现出对比同类方法的优越性与更好的识别效果。Due to factors such as viewpoint changes,complex environments,and pose differences under multiple cameras,the images of the same vehicle in different scenes show huge appearance ambiguity,which brings challenges to vehicle identity matching.In order to solve this problem,a feature robustness enhancement method for vehicle recognition is proposed under the transformer framework.Based on the fact that the structural information of the vehicle is invariant under multiple cameras,a module for enhancing structural information guided by contour features is designed,and a structural feature perception loss is proposed to promote the fusion of structural information in the model.The attribute information of the vehicle is embedded into the transformer framework as a vector,which further alleviates the influence of vehicle pose changes under multiple viewpoints.Experimental results show that the proposed method exhibits superiority and better recognition effect compared with similar methods on VeRi-776 and VehicleID datasets.

关 键 词：车辆识别 双流网络 结构信息 矢量嵌入 域不变特征提取 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]