文章摘要
张立国,李佳庆,赵嘉士,耿星硕,章玉鹏.基于改进YOLOv4的灾后人员检测算法[J].高技术通讯(中文),2023,33(7):742~749
基于改进YOLOv4的灾后人员检测算法
Post-disaster personnel detection algorithm based on improved YOLOv4
  
DOI:10.3772/ j. issn.1002-0470.2023.07.008
中文关键词: 灾后救援; 目标检测; 改进YOLOv4; Mobilenetv1; K-means++
英文关键词: post-disaster relief, object detection, improved YOLOv4, MobileNetv1, K-means++
基金项目:
作者单位
张立国 (燕山大学电气工程学院秦皇岛 066004) (河北省测试计量技术及仪器重点实验室秦皇岛 066004) 
李佳庆  
赵嘉士  
耿星硕  
章玉鹏  
摘要点击次数: 470
全文下载次数: 403
中文摘要:
      针对地震等灾后环境复杂、救援机器人对救援目标识别实时性和准确度要求较高的问题,提出一种基于改进YOLOv4的目标检测模型。该算法将YOLOv4网络中的主干特征提取网络替换成MobileNetv1模型以增强特征复用,同时缩减网络参数量,提高运行速度;通过K-means++算法进行锚点维度聚类以适应灾后人员检测,提升算法精度。并且针对灾后人员检测数据集缺乏问题,贡献相应的数据集。实验结果表明,改进的网络与YOLOv4相比在保证模型精度的前提下帧率提升了约92%,权重文件大小变为原来的20.73%,满足了灾后救援机器人目标检测实时性和准确性的需求,对于灾后人员检测场景有一定的借鉴意义。
英文摘要:
      A target detection model based on improved YOLOv4 is proposed for the complex post-disaster environments such as earthquakes and rescue robots with high requirements for real-time and accurate recognition of rescue targets.This algorithm replaces the backbone feature extraction network in the YOLOv4 network with the MobileNetv1 model to enhance feature reuse, while reducing the number of network parameters and improving operational speed. Anchor dimension clustering by K-means++ algorithm to adapt to post-disaster personnel detection and improve algorithm accuracy. And it contributes to addressing the problem of missing datasets for post-disaster personnel detection. Experimental results show that the improved network improves the frame rate by about 92% compared with YOLOv4 in guaranteed model accuracy, and the weight file size becomes 20.73% of the original. The demand for real-time and accurate target detection of post-disaster rescue robots is satisfied, and this study has certain implications for post-disaster personnel detection scenarios.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮