基于FRL-Net的高鲁棒性多尺度小样本轨道入侵异物检测方法研究
Research on the high robust multi-scale few-shot railway intrusion obstacles detection method based on FRL-Net
赵宗扬 1康杰虎 1吴斌 1叶涛 2张振1
作者信息
- 1. 天津大学精密测量技术及仪器全国重点实验室 天津 300072
- 2. 中国矿业大学(北京) 机电与信息工程学院 北京 100083
- 折叠
摘要
针对轨道入侵异物严重威胁行车安全,而基于广义深度学习的目标检测方法无法打破大数据驱动的训练壁垒、小样本目标检测方法在复杂轨道环境中对多尺度入侵异物检测能力差、鲁棒性低等问题,本文提出了一种高鲁棒性多尺度小样本轨道入侵异物检测模型.该模型采用元学习策略,通过多尺度小样本入侵异物特征提取模块增强模型对于不同尺度小样本异物特征的表达能力.使用轨道入侵异物元特征精准重加权模块对小样本异物的元特征进行精准优化.提出小样本轨道入侵异物检测优化模块进一步提升模型的检测性能.实验结果表明,该模型在 7-way 30-shot的小样本轨道异物检测任务中的平均检测精度为 81.8%,比FSRW高 3.2%,更适合在实际轨道环境中检测多尺度小样本入侵异物.
Abstract
Aiming at the serious threat to train safety posed by the railway intrusion obstacles,while the general object detection methods based on deep learning struggle to break the barrier of data-driven training,the few-shot object detection methods have weak detection ability and low robustness for multi-scale obstacles in complex railway environments,this paper presents a high robust multi-scale few-shot railway intrusion obstacles detection model(FRL-Net).The model utilizes the meta-learning strategy to capture rich feature information by designing the multi-scale few-shot obstacle feature extraction module,which can enhance the model's ability to express the features of few-sample objects at different scales.The precise reweighting module is used for optimizing the meta-feature at different scales,and the few-shot railway obstacle detection optimization module is proposed to further enhance the few-shot railway obstacle detection performance of the model.The experimental results show that the proposed model achieves the mAP of 81.8%in the 7-way 30-shot few-shot railway obstacle detection task,which is 3.2%higher than that of FSRW.It is more suitable for detecting few-shot multi-scale railway obstacles in actual railway environments.
关键词
轨道入侵异物/元学习/小样本目标检测/多尺度/深度学习Key words
railway intrusion obstacles/meta learning/few-shot object detection/multi-scale/deep learning引用本文复制引用
基金项目
天津市交通运输科技发展计划(2022-09)
北京市自然科学基金(L221018)
光纤传感与系统北京实验室开放基金(GXKF2022001)
天津大学自主创新基金(2023XHX-0019)
出版年
2024
NETL
NSTL
万方数据