Scholarlink

<p>Vehicle detection is an important technology to ensure the safe driving of autonomous vehicles in traffic flow. Some existing lightweight vehicle detection models exhibit issues such as missed detections and incorrect detections. This paper proposed a fast and accurate front vehicle detection model. Firstly, the feature extraction network was replaced by Ghost-HGNet, designed based on PPHGNetv2, which reduced the parameters and calculations of the model and enhanced the ability of the model to extract features. Secondly, the Efficient Multi-Scale Attention (EMA) mechanism was introduced into the SPPF module to enhance the interaction of important feature information, and to improve the model’s performance on vehicle detection in the case of severe occlusion. Thirdly, the Half Receptive-Field Attention Convolution Network (HRACN) module was introduced, based on the Receptive-Field Attention Convolution (RFAConv) and C2f module. This model effectively extracted the receptive field features and enhanced the model’s ability to extract vehicle target information under different backgrounds. Compared with YOLOv8n, the improved model increased the Recall by 2.7% and the mAP0.5 by 1.4% on the KITTI dataset, reduced the number of parameters by 20%, and reduced flops by 10%. The FPS of the improved model reached 122.5, which met the requirements of real-time detection. Additionally, the improved model demonstrated good generalization performance on the BDD100K dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A lightweight model for autonomous vehicle to detect front vehicles in traffic flow

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Xuejing Du,
  • Shixin Liu

摘要

Vehicle detection is an important technology to ensure the safe driving of autonomous vehicles in traffic flow. Some existing lightweight vehicle detection models exhibit issues such as missed detections and incorrect detections. This paper proposed a fast and accurate front vehicle detection model. Firstly, the feature extraction network was replaced by Ghost-HGNet, designed based on PPHGNetv2, which reduced the parameters and calculations of the model and enhanced the ability of the model to extract features. Secondly, the Efficient Multi-Scale Attention (EMA) mechanism was introduced into the SPPF module to enhance the interaction of important feature information, and to improve the model’s performance on vehicle detection in the case of severe occlusion. Thirdly, the Half Receptive-Field Attention Convolution Network (HRACN) module was introduced, based on the Receptive-Field Attention Convolution (RFAConv) and C2f module. This model effectively extracted the receptive field features and enhanced the model’s ability to extract vehicle target information under different backgrounds. Compared with YOLOv8n, the improved model increased the Recall by 2.7% and the mAP0.5 by 1.4% on the KITTI dataset, reduced the number of parameters by 20%, and reduced flops by 10%. The FPS of the improved model reached 122.5, which met the requirements of real-time detection. Additionally, the improved model demonstrated good generalization performance on the BDD100K dataset.