期刊文献+

雷达占据栅格图车辆目标的轻量化YOLOv5s方法研究

Lightweight YOLOv5s Method for Vehicle Targets in Radar Occupancy Grid Map
下载PDF
导出
摘要 毫米波雷达传感器可穿透火障、烟瘴及高温,是特长隧道等封闭空间灾中车辆目标检测的主要信息来源。基于雷达的同时定位与构图技术(Simultaneous Localization and Mapping,SLAM)可对受灾的静止车辆及环境进行占据栅格成像,并在低成本边缘计算平台上进行高精度实时车辆目标检测。现有YOLOv5s目标检测网络模型过于复杂,且训练模型针对光学数据,与雷达占据栅格地图(Radar Occupancy Grid Map,ROGM)中的车辆目标特征存在差异。针对上述问题提出了YOLOv5s-MDS轻量化网络模型,以雷达图像特征为约束优化网络结构,参数量相对原YOLOv5s模型降低42.3%。实验表明:该网络模型在大规模VOC光学图像预训练模型辅助下,利用雷达小样本数据微调可快速收敛。 The millimeter wave radar sensor can penetrate the fire barrier,smoke and high temperature,and is the main source of information for vehicle target detection in closed space disasters such as extra-long tunnels.Radar based simultaneous localization and mapping(SLAM)technology can conduct occupancy grid imaging of the affected static vehicles and environment,and conduct high-precision real-time vehicle target detection on the low-cost edge computing platform.The existing YOLOv5s target detection network model is too complex,and the training model is based on optical data,which is different from the vehicle target features in the radar occupancy grid map(ROGM).In this paper,YOLOv5s-MDS lightweight network model is proposed to solve the above problems.The network structure is optimized with radar image characteristics as constraints,and the parameter quantity is reduced by 42.3%compared with the original YOLOv5s model.The experiment in this paper shows that the network model can converge quickly by fine-tuning the radar small sample data with the aid of large-scale VOC optical image pre-training model.
出处 《工业控制计算机》 2023年第9期1-3,共3页 Industrial Control Computer
基金 国家自然科学基金重点项目“城市重大基础设施近景成像雷达动态安全监测成像模型与方法研究”(62131001)。
关键词 车辆目标检测 ROGM YOLOv5s 轻量化 MobileNetV3 注意力机制 vehicle target detection ROGM YOLOv5s lightweight MobileNetV3 attention mechanism
  • 相关文献

参考文献2

二级参考文献14

共引文献238

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部