刘洪龙,李向阳,徐正华,卢朝晖.基于SqueezeNet和YOLOv2的交通违法证据评价[J].南华大学学报(自然科学版),2022,(1):80~87.[LIU Honglong,LI Xiangyang,XU Zhenghua,LU Zhaohui.Traffic Violation Evidence Evaluation Based on YOLOv2 and SqueezeNet[J].Journal of University of South China(Science and Technology),2022,(1):80~87.] |
基于SqueezeNet和YOLOv2的交通违法证据评价 |
Traffic Violation Evidence Evaluation Based on YOLOv2 and SqueezeNet |
投稿时间:2021-09-14 |
DOI:10.19431/j.cnki.1673-0062.2022.01.012 |
中文关键词: 交通违法 图像识别 检测卷积网络 图像质量 迁移学习 |
英文关键词:traffic violation image recognition detection of a convolutional network image quality transfer learning |
基金项目:湖南省社会发展领域重点研发项目(2019SK2011);湖南省教育厅重点项目(20A440);国防科工局“十三五”技术基础科研项目(403C001);国防科工局“十三五”技术基础科研项目(2018年)(403B01) |
|
摘要点击次数: 562 |
全文下载次数: 336 |
中文摘要: |
针对交通监控反向抓拍交通违法图像预判率高的问题,提出了一种基于迁移学习的多尺度交通违法证据评价方法。构建了以SqueezeNet为特征提取层、YOLOv2为目标检测层融合高分辨率细粒度特征的检测网络。通过卷积神经网络算法训练该模型学习抓拍车辆图像特征,识别图像中唯一交通违法车辆,再次训练识别驾驶员所在中心区域。在保证特征识别提取能力不变的条件下,采用迁移学习的方式重新训练SqueezeNet,对驾驶员中心区域图像进行好坏二分类,将成像清晰的交通违法图像提交给人工审核。实验结果表明此方法将违法车辆检测准确率提升到99.3%,驾驶员所在关键区域检测准确率提升到96.3%,图像质量评价准确率提升到92.6%,大幅降低了人工审核工作量。 |
英文摘要: |
Aiming at the problem of the high predictive rate of traffic violation images captured by reverse traffic monitoring, a multi-scale traffic violation evidence evaluation method based on migration learning is proposed. A detection network is constructed that uses SqueezeNet as the feature extraction layer and YOLOv2 as the target detection layer to fuse high-resolution fine-grained features. Through the convolutional neural network algorithm, the model is trained to learn the characteristics of the captured vehicle image, identify the only illegal traffic vehicle in the image, and retrain to identify the central area where the driver is located. Under the condition that the ability of feature recognition and extraction remains unchanged, SqueezeNet is retrained by transfer learning to classify the image of the driver's central area as good or bad, and the clear image of traffic violations is submitted to manual review. Experimental results show that this method improves the detection accuracy of illegal vehicles to 99.3%, the detection accuracy of key areas where the driver is located to 96.3%, and the image quality evaluation accuracy to 92.6%, which greatly reduces the workload of manual review. |
查看全文 查看/发表评论 下载PDF阅读器 |
关闭 |
|
|
|