摘要
基于卷积神经网络的目标检测在智能交通领域有着重要的应用,但存在复杂网络模型计算速度慢、简单网络模型精准度低两种问题.针对此问题,本文提出了基于Lite-YOLOv3的行人与车辆检测方法,该方法基于Tiny-YOLOv3网络模型进行改进.首先,本文采用卷积代替下采样方案解决Tiny-YOLOv3网络特征提取损失问题.然后其骨干层采用改进的瓶颈块(BottleneckBlock)对前一层网络特征图进行降维、连接输入输出特征图,使得网络参数量大幅下降、防止网络退化.其预测层采用改进后的深度可分离卷积块(Depthwise Separable Convolution),分离深度卷积和点卷积可以有效降低网络运算成本,加快网络运算速度.Lite-YOLOv3相较于Tiny-YOLOv3网络的运算速度提升了27.27%,mAP提高了9.07%.
Object detection based on convolutional neural networks has important applications in the field of intelligent transportation, but there are two problems: slow calculation speed of complex network models and low accuracy of simple network models.To solve this problem, this paper proposes a pedestrian and vehicle detection method based on Lite-YOLOv3,which is improved based on the Tiny-YOLOv3 network model.First of all, this paper uses convolution instead of downsampling to solve the loss problem of Tiny-YOLOv3 network feature extraction.Then its backbone layer uses an improved Bottleneck Block to reduce the dimensionality of the network feature map of the previous layer and connect the input and output feature maps, so that the amount of network parameters is greatly reduced and network degradation is prevented.Its prediction layer adopts an improved Depthwise Separable Convolution(Depthwise Separable Convolution).Separating deep convolution and point convolution can effectively reduce network operation costs and speed up network operations.Compared with the Tiny-YOLOv3 network, Lite-YOLOv3 has a 27.27% increase in computing speed, and mean average precision has increased by 9.07%.
作者
涂媛雅
汤国放
张建勋
TU Yuan-ya;TANG Guo-fang;ZHANG Jian-xun(College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第1期211-217,共7页
Journal of Chinese Computer Systems
基金
重庆市教育委员会科学技术重点研究项目(KJZD-K201801901)资助。
作者简介
涂媛雅,女,1997年生,硕士研究生,CCF会员,研究方向为计算机图形学、目标检测;汤国放,男,1995年生,硕士研究生,CCF会员,研究方向为计算机图形学、目标检测;通讯作者:张建勋,男,1971年生,博士,教授,CCF会员,研究方向为计算机图形学、目标检测,E-mail:timj3ly@gmail.com。
