基于直方图特征蒸馏的密集交通目标检测  

Dense Traffic Object Detection Based on Histogram Feature Distillation

作　　者：张亿鸿 钟铭恩 谭佳威 范康 李正峰 Zhang Yihong;Zhong Mingen;Tan Jiawei;Fan Kang;Li Zhengfeng(School of Mechanical and Automotive Engineering,Xiamen University of Technology,Xiamen 361024;School of Aerospace Engineering,Xiamen University,Xiamen 361005)

机构地区：[1]厦门理工学院机械与汽车工程学院,厦门361024 [2]厦门大学航空航天学院,厦门361005

出　　处：《汽车工程》2025年第4期636-644,679,共10页Automotive Engineering

基　　金：福建省自然科学基金(2023J011439)资助。

摘　　要：密集交通场景下的多类交通参与者目标检测仍是一项颇具挑战的视觉任务,对于交通管理和安全至关重要。为此,针对密集交通参与者的局部遮挡和小尺度特点,提出一种深度神经网络检测算法DSODet。首先采用轻量化的CSPDarkNet网络提取交通图像特征;然后设计了多尺度特征融合上采样模块以增强对难检测目标的表达能力;随后增加高分辨率检测分支来提升对小尺度目标的检测能力;最后,提出直方图特征蒸馏训练方法,通过最小化教师模型与学生模型相同层特征直方图的交集比,来有效引导学生模型训练,实现参数优化与轻量化。实验结果表明,DSODet对交通参与者的平均检测精度为66.9%,对局部遮挡的小尺度目标为13.0%,均超越现有主流算法,模型参数量仅为2.9 M,体现了对边缘设备的友好性。相关代码将在https://github.com/XMUT-Vsion-Lab分享。Multi-class traffic participant detection in dense traffic scenarios remains a challenging visual task,which is crucial for traffic management and safety.To address this,a deep neural network-based detection algorithm,DSODet,is proposed to handle the challenges of partial occlusion and small-scale targets in dense traffic environment.Firstly,a lightweight CSPDarkNet network is used to extract features from traffic images.Then,a multi-scale feature fusion upsampling module is designed to enhance the representation capability for hard-to-detect targets.Next,a high-resolution detection branch is incorporated to improve detection accuracy for small-scale targets.Finally,a histogram feature distillation training method is proposed,which effectively guides the student model′s training by minimizing the intersection ratio of feature histograms between the teacher and student models at corresponding layers,thus enabling parameter optimization and model compression.The experimental results show that DSODet achieves an average detection accuracy of 66.9% for traffic participants and 13.0% for small targets with partial occlusion,outperforming current state-of-the-art algorithms.The model contains only 2.9 M parameters,demonstrating its friendliness for edge device.The related code will be shared at https://github.com/XMUT-VsionLab.

关 键 词：目标检测 密集交通 小尺度目标 局部遮挡 直方图特征蒸馏 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]