摘要
基于智能视频分析的轨道线路环境入侵物自主识别是保障轨道交通运营安全的关键技术之一。然而基于神经网络的高精度目标检测模型严重依赖算力,部署成本高,很难普及运用。为此,提出一种改进yolov4-tiny的轻量级网络模型。在网络主干,通过融合跨阶段结构和通道混洗策略,提出CSPShuffleNet结构,加快网络推理;在网络颈部,引入多头注意力机制,增强网络目标定位能力;在网络头部,使用深度可分离卷积替换传统卷积,进一步压缩网络参数量。基于铁路异物数据集的实验结果表明:相比于原始yolov4-tiny,本模型的均值平均精度最大提高1.4%,参数量减少49.9%,模型容量减少55.4%。验证了本模型对于固定平台和移动平台检测系统的普适性,从而为铁路安全保障提供决策支持。
The automatic identification of intrusions on the track under the railway environment based on intelligent video analysis is one of the key technologies to ensure the safety of train operations.The high precision object detection model based on neural networks heavily relies on computing power,which leads to high deployment costs and difficulty for wide applications.To deal with this,this paper proposed a lightweight model based on improved yolov4-tiny network.In network’s backbone,by integrating the cross stage partial architecture and channel shuffling strategy,a CSPShuffleNet module was proposed to accelerate the network inference.In network’s neck,the multi-head attention mechanism was introduced to enhance the network ability of object location.In network’s head,the depth-wise separable convolution was also used to replace the vanilla convolutional layer to further reduce the number of parameters.The experimental results based on the railway intrusion datasets show that when compared with the original yolov4-tiny,the proposed model shows a maximum improvement of mean average precision by 1.4%,a reduction of the number of parameters by 49.9%,and a reduction of the model size by 55.4%.This paper verifies the universality of the model applied in the fixed platform and mobile platform detection system,which can provide decision support for the railway security.
作者
管岭
贾利民
谢征宇
GUAN Ling;JIA Limin;XIE Zhengyu(School of Traffic and Transportation,Beijing Jiaotong University,Beijing 100044,China;State Key Laboratory of Rail Traffic Control and Safety,Beijing Jiaotong University,Beijing 100044,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2023年第5期72-81,共10页
Journal of the China Railway Society
基金
北京交通大学基本科研业务费(2019YJS095)。
关键词
异物入侵检测
轻量化神经网络
深度可分离卷积
通道混洗
多头注意力机制
intrusion detection
lightweight neural network
depth-wise separable convolution
channel shuffle
multi-head attention mechanism
作者简介
第一作者:管岭(1993—),男,安徽肥东人,博士研究生。E-mail:18114020@bjtu.edu.cn;通信作者:谢征宇(1983—),女,浙江天台人,副教授,博士。E-mail:xiezhengyu@bjtu.edu.cn。
