摘要
随着快递及外卖业的兴起,以电动车和摩托车为代表的双轮车数量激增,交通事故频发。由于双轮车数量庞大,管理将耗费大量警力,而该研究能够大大释放警力。针对目标检测模型中高分辨率特征层计算耗时的问题,提出了采用稀疏卷积的级联双轮车头盔目标检测算法,有效提升模型性能,速度提高了33.3%。此外针对行人以及自行车驾乘人员带来的未戴头盔误判问题,采用多尺度空洞卷积,通过引入上下文信息,可以有效减少此类误判,精度提升2.2%。最后标注并开源了交通道路场景下的双轮车头盔数据集TWHD,以验证算法性能。
With the rise of express delivery industry,the number of two wheeler represented by electric vehicles and motorcycles has increased sharply,and traffic accidents occur frequently.Due to the large number of two wheeled vehicles,the management will consume a lot of police force,and this study can greatly release the police force.To solve the time-consuming problem of high-resolution feature layer calculation in the object detection algorithm,this paper proposes a Cascade two wheeler helmet detection algorithm using sparse convolution,which improves the speed by 33.3%.In addition,for the misjudgment of pedestrians and cyclists without helmets,multi-scale dilated convolution is adopted.By introducing context information,such misjudgment can be effectively reduced and the accuracy can be improved by 2.2%.Finally,we annotate and open source TWHD dataset to verify the performance of the algorithm.
作者
李丹峰
LI Dan-feng(School of Computer Science,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《信息技术》
2024年第3期56-62,69,共8页
Information Technology
基金
浙江省大学生科技创新活动计划暨新苗人才计划项目(2021R407026)。
关键词
深度学习
双轮车头盔目标检测
小目标检测
稀疏卷积
多尺度空洞卷积
deep learning
two wheeler helmet detection
small object detection
sparse convolution
multi-scale dilated convolution