Scholarlink

<p>The purpose of this study is to solve the problem of how to balance the real-time performance and detection accuracy when You Only Look Once version 8 (YOLOv8) model is applied to the navigation task of mobile robots with limited resources. The main limitations stem from its high computational cost and insufficient detection ability for small and occluded targets in complex environments. To overcome these problems, a coordinate attention YOLOv8 (GCA-YOLOv8) model based on GhostNet is proposed. Lightweight GhostNetV2 is introduced to replace the original backbone structure of YOLOv8. This modification significantly reduces the model parameters and the computational complexity of one billion Floating-Point Operations Per Second (GFLOPs). Its embedded, decoupled and fully connected attention mechanism effectively captures the long-distance dependence in the feature graph. In addition, the original path aggregation network is replaced by a two-way feature pyramid network. Through adaptive multi-scale feature fusion with learnable weights, the robustness of object detection at different scales is improved. In addition, the coordination attention mechanism is embedded in the feature fusion layer of the network. By encoding the directional awareness, location-sensitive spatial information and the dependencies between channels at the same time, this mechanism improves the positioning accuracy and guides the model to focus on key target areas. The original loss function is further replaced by Wise-Intersection over Union. Its dynamic non-monotonic focusing mechanism optimizes the bounding box regression, reduces the adverse effects of low-quality samples in the training process, accelerates the convergence speed of the model and improves the detection accuracy. To verify the effectiveness of the proposed method, comparative experiments and ablation research are carried out on the Common Objects in Context dataset. The results show that compared with YOLOv8n, GCA-YOLOv8 reduces parameters by 20.8%, GFLOPs by 25.3%, and mAP@0.5 increases by 2.1%. The real-time inference speed of 51.2 frames per second is realized on the embedded platform of NVIDIA Jetson AGX Xavier. On the whole, this study focuses on the edge deployment scene of mobile robots, and makes hierarchical arrangement and joint optimization of GhostNetV2, Coordinate Attention (CA) and WIoU, and finally forms a lightweight customization scheme for robot visual guidance.</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

Optimized YOLOv8-based target detection model for robot visual guidance

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月21日

查看全文
  • Hongwei Li

摘要

The purpose of this study is to solve the problem of how to balance the real-time performance and detection accuracy when You Only Look Once version 8 (YOLOv8) model is applied to the navigation task of mobile robots with limited resources. The main limitations stem from its high computational cost and insufficient detection ability for small and occluded targets in complex environments. To overcome these problems, a coordinate attention YOLOv8 (GCA-YOLOv8) model based on GhostNet is proposed. Lightweight GhostNetV2 is introduced to replace the original backbone structure of YOLOv8. This modification significantly reduces the model parameters and the computational complexity of one billion Floating-Point Operations Per Second (GFLOPs). Its embedded, decoupled and fully connected attention mechanism effectively captures the long-distance dependence in the feature graph. In addition, the original path aggregation network is replaced by a two-way feature pyramid network. Through adaptive multi-scale feature fusion with learnable weights, the robustness of object detection at different scales is improved. In addition, the coordination attention mechanism is embedded in the feature fusion layer of the network. By encoding the directional awareness, location-sensitive spatial information and the dependencies between channels at the same time, this mechanism improves the positioning accuracy and guides the model to focus on key target areas. The original loss function is further replaced by Wise-Intersection over Union. Its dynamic non-monotonic focusing mechanism optimizes the bounding box regression, reduces the adverse effects of low-quality samples in the training process, accelerates the convergence speed of the model and improves the detection accuracy. To verify the effectiveness of the proposed method, comparative experiments and ablation research are carried out on the Common Objects in Context dataset. The results show that compared with YOLOv8n, GCA-YOLOv8 reduces parameters by 20.8%, GFLOPs by 25.3%, and mAP@0.5 increases by 2.1%. The real-time inference speed of 51.2 frames per second is realized on the embedded platform of NVIDIA Jetson AGX Xavier. On the whole, this study focuses on the edge deployment scene of mobile robots, and makes hierarchical arrangement and joint optimization of GhostNetV2, Coordinate Attention (CA) and WIoU, and finally forms a lightweight customization scheme for robot visual guidance.