文献类型：期刊文献

中文题名：结合多尺度特征的改进YOLOv2车辆实时检测算法

英文题名：Improved YOLOv2 vehicle real-time detection algorithm combined with multi-scale features

作者：金宇尘[1];罗娜[1]

机构：[1]华东理工大学化工过程先进控制和优化技术教育部重点实验室,上海200237

年份：2019

卷号：40

期号：5

起止页码：1457

中文期刊名：计算机工程与设计

外文期刊名：Computer Engineering and Design

收录：CSTPCD;;北大核心:【北大核心2017】；

基金：国家自然科学基金项目(61403140)

语种：中文

中文关键词：无人驾驶系统;车辆目标;实时检测;深度学习;难样本生成

外文关键词：automatic driving system;vehicle target;real-time detection;deep learning;hard sample generation

摘要：为解决车辆目标检测过程中较小车辆及被遮挡车辆容易被漏检、错检的问题,基于深度学习先进算法YOLOv2,提出一种结合多尺度特征的改进车辆目标实时检测方法。通过融合多尺度特征层信息建立额外检测特征层对大小不同的车辆进行检测;构建自适应损失函数作为置信度目标函数解决正负样本数量不平衡问题,提高置信度预测准确性;开发一种新的难样本生成方法对网络进行训练,减小错检发生的概率。实验结果表明,该方法在运行速度满足实时检测要求的情况下,能显著降低车辆目标漏检、错检率,平均准确率提高9%以上。
To solve missed and false detection of smaller vehicles and obscured vehicles in detection process,based on the advanced algorithm of deep learning,YOLOv2,a vehicle real-time detection method combined with multi-scale features was proposed.An additional detection feature layer was set up through the combination of multi-feature layer information to detect vehicles in different sizes.The adaptive loss function was constructed as the confidence function to solve the imbalance between positive and negative samples,which improved the accuracy of confidence prediction.A new hard sample generation method was developed to train the network and reduce the possibility of false detection.Experimental results show that the speed of the proposed method meets the requirement of real-time detection,and it can significantly reduce the missed and false detection rates and increase the average precision by more than 9%.