基于高阶空间特征聚合的车型识别算法  

Vehicle recognition algorithm based on high-order spatial feature aggregation

作　　者：杨潞霞 薛映昭 张红瑞 马永杰[3] Yang Luxia;Xue Yingzhao;Zhang Hongrui;Ma Yongjie(School of Computer Science and Technology,Taiyuan Normal University,Jinzhong 030619,China;Shanxi Provincial Key Laboratory of Intelligent Optimization Computing and Blockchain Technology,Jinzhong 030619,China;School of Physics and Electronic Engineering,Northwest Normal University,Lanzhou 730070,China)

机构地区：[1]太原师范学院计算机科学与技术学院,晋中030619 [2]智能优化计算与区块链技术山西省重点实验室,晋中030619 [3]西北师范大学物理与电子工程学院,兰州730070

出　　处：《电子测量技术》2025年第4期169-180,共12页Electronic Measurement Technology

基　　金：国家自然科学基金(62066041);山西省重点研发计划(202102010101008);山西省高等学校科技创新项目(2024L295)资助。

摘　　要：针对复杂交通场景下车型目标密集、遮挡而造成的车型识别精度低的问题,提出基于高阶空间特征聚合的车型识别算法。首先,在特征提取的下采样阶段,设计了HSIDM模块,实现更深层次的特征聚合,减少细小信息损失。其次,在特征融合部分设计了DMFAM模块,动态调整各尺度特征的权重,获取多尺度的上下文信息,以增强模型对多样化特征的适应能力。然后,设计解耦REL-Head检测头,将分类和回归任务拆解,避免任务混杂,增强局部特征的学习能力与抗干扰能力。最后,将本文模型部署到边缘设备进行测试。实验结果显示,本文算法在复杂度交通场景数据集BIT-Vehicle和UA-DETRAC上,mAP相较于原模型分别提升了0.7%和3.9%,并在边缘设备上可以流畅运行,具有较好的识别效果。表明所提出的方法能够有效提高车型识别的精度并应用于受限设备。Addressing the issue of low vehicle recognition accuracy stemming from dense vehicle targets and occlusions in complex traffic scenarios,a vehicle recognition algorithm utilizing high-order spatial feature aggregation is proposed.Initially,during the downsampling phase of feature extraction,the HSIDM module is devised to facilitate deeper feature aggregation and minimize the loss of fine details.Subsequently,within the feature fusion component,the DMFAM module is introduced to dynamically adjust the weights of features across various scales,thereby acquiring multi-scale contextual information and bolstering the model′s adaptability to diverse features.Following this,a decoupled REL-Head detector is formulated to disentangle classification and regression tasks,preventing task mixing and enhancing the learning capability and interference resistance of local features.Ultimately,the model presented in this paper is deployed on edge devices for testing.Experimental outcomes reveal that on the complex traffic scene datasets BIT-Vehicle and UA-DETRAC,the mean average precision(mAP)of our algorithm has improved by 0.7%and 3.9%respectively compared to the original model.Additionally,it operates seamlessly on edge devices,demonstrating effective recognition capabilities.This indicates that the proposed approach can effectively enhance the precision of vehicle identification and is suitable for use on constrained devices.

关 键 词：车型识别 高阶空间交互 动态多尺度特征聚合 解耦检测头 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN0[自动化与计算机技术—计算机科学与技术]