Scholarlink

<p>Existing target detection algorithms have limitations in complex mine environments like low illumination, small targets, background interference, occlusion, and motion blur. Also, their complex network structures and large parameter volumes can't meet real-time detection needs of edge devices. Thus, a lightweight network-based multi-target detection method for mine driverless rail locomotive driving areas was proposed. A dataset of seven target images (electric locomotives, miners, etc.) in five scenarios (normal &amp; low illumination, etc.) was constructed. Based on YOLOv5s, improvements were made: adding a small target detection layer to enhance small target detection; using the GhostBottleNeck module to replace BottleNeck in C3 to build C3Ghost, reducing calculation and parameters and compensating for the added layer; introducing the SimAM attention mechanism to focus on targets and suppress interference; replacing CIoU with SIoU loss function to speed up convergence. Experimental results show the proposed lightweight network cuts parameters by 12.3%, boosts mAP by 1.7%, and is more suitable for multi-object detection in mine rail locomotive driving areas.</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

Multi-target detection method for driving area of mine driverless rail locomotives based on lightweight network

  • 期刊论文

  • 发表日期：

    2025年2月10日

查看全文
  • Wenshan Wang,
  • Kun Hu,
  • Hao Jiang

摘要

Existing target detection algorithms have limitations in complex mine environments like low illumination, small targets, background interference, occlusion, and motion blur. Also, their complex network structures and large parameter volumes can't meet real-time detection needs of edge devices. Thus, a lightweight network-based multi-target detection method for mine driverless rail locomotive driving areas was proposed. A dataset of seven target images (electric locomotives, miners, etc.) in five scenarios (normal & low illumination, etc.) was constructed. Based on YOLOv5s, improvements were made: adding a small target detection layer to enhance small target detection; using the GhostBottleNeck module to replace BottleNeck in C3 to build C3Ghost, reducing calculation and parameters and compensating for the added layer; introducing the SimAM attention mechanism to focus on targets and suppress interference; replacing CIoU with SIoU loss function to speed up convergence. Experimental results show the proposed lightweight network cuts parameters by 12.3%, boosts mAP by 1.7%, and is more suitable for multi-object detection in mine rail locomotive driving areas.