摘要
针对铁路施工环境较复杂,安全穿戴目标较小难以检测,边缘计算设备资源有限的问题,提出一种基于改进RT-DETR的铁路施工场景下人员安全穿戴检测模型。首先,引用轻量级EfficientViT作为特征提取网络,通过级联分组注意力,解决多头自注意力计算冗余问题,提高注意力头的多样性。其次,采用HWD-ADown下采样模块,应用Haar小波变换保留更多细节信息来改善错检问题,通过将特征图切分再进行卷积的方式减少卷积操作的参数量,进一步降低模型复杂度,精度维持原来相近水平。最后,设计一种新的损失函数Inner-DIoU,在加速边界框回归速度的同时提高模型检测的泛化能力。实验结果表明,改进模型精确率为92.6%,召回率为84.4%,平均精度均值为90%,与基准模型相比分别提高2.7%、2.1%和3%;模型大小为19.9 MB,参数量为985.6万个,GFLOPs为25.5,与基准模型相比分别降低48.4%、50.4%和55.4%;FPS为94.3,提高了34.7%。提出的模型能够满足铁路施工场景下对检测精度和轻量化的需求。
Wearing reflective clothing and helmet is an important guarantee for the life safety of workers in the construction scenes.In order to solve the problems of complex railway construction environment,small safety wear targets that are difficult to detect,and limited edge computing equipment resources,an improved RT-DETR personnel safety wear detection model for railway construction scenes was proposed.Firstly,the lightweight network EfficientViT was cited as the model backbone feature extraction network to solve the computational redundancy problem of multi-head self-attention through cascaded group attention,and improve the diversity of attention heads.Secondly,the HWD-ADown module was used to replace the subsampling module in the original RT-DETR.The Haar wavelet transform was applied to retain more details to improve the error detection problem.The number of parameters in subsampling was reduced by segmental convolution of feature graphs,further reducing the model complexity while maintaining the accuracy close to the original level.Finally,a new loss function was designed,which combined the Inner-IoU idea with DIoU to accelerate the bounding box regression speed and improve the generalization ability of model detection.The results show that the improved model achieves an accuracy rate of 92.6%,a recall rate of 84.4%,and an average accuracy of 90%,which are 2.7%,2.1%and 3%higher than the benchmark model,respectively.The results show that the improved model achieves an accuracy rate of 92.6%,a recall rate of 84.4%,and an average accuracy of 90%,which are 2.7%,2.1%and 3%higher than the benchmark model,respectively.The size of the model is 19.9 MB,with 9.856 million parameters,and 25.5 GFLOPs,which are 48.4%,50.4%and 55.4%lower than the benchmark model,respectively.FPS(frames per second)is 94.3,an improvement of 34.7%.The improved model can meet the requirements of detection accuracy and lightweight in railway construction scenarios.
作者
冯爽
王万齐
杨文
胡昊
FENG Shuang;WANG Wanqi;YANG Wen;HU Hao(Postgraduate Department,China Academy of Railway Sciences,Beijing 100081,China;Institute of Computing Technology,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China)
出处
《铁道学报》
北大核心
2025年第2期92-101,共10页
Journal of the China Railway Society
基金
中国国家铁路集团有限公司科技研究开发计划(K2023T003)
第二十七届中国科协年会学术论文。
作者简介
第一作者:冯爽(1999-),女,山东聊城人,硕士。E-mail:fengshuangba@163.com;通信作者:王万齐(1978-),男,甘肃庆阳人,研究员,博士。E-mail:wangwanqi@rails.cn。
