Parallel RepConv network: Efficient vineyard obstacle detection with adaptability to multi-illumination conditions

文献类型： 外文期刊

第一作者： Cui, Xuezhi

作者： Cui, Xuezhi;Zhu, Licheng;Zhao, Bo;Wang, Ruixue;Han, Zhenhao;Zhang, Weipeng;Dong, Lizhong

作者机构：

关键词： Obstacle detection; Parallel RepConv network (PRCN); Vineyard environment; Balance

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 230 卷

页码：

收录情况： SCI

摘要： Obstacle detection is crucial for autonomous navigation operating in orchard environments. This study introduces the Parallel RepConv Network (PRCN), a novel and efficient convolutional neural network (CNN) designed specifically for vineyard obstacle detection. PRCN balances speed and accuracy through several key innovations. Its backbone utilizes the Parallel RepConv (PRC) operation block, composed of two distinct RepConv modules with residual connections for enhanced feature extraction. Multi-scale features from the backbone are then fused using TriangleNet, a lightweight and efficient feature fusion network employing two different fusion methods, its compact design, requiring only five operations, contributes significantly to both detection accuracy and fast inference. To improve robustness to varying lighting conditions, a novel data augmentation technique called "Transition" generates composite images representing diverse illumination throughout the day. Furthermore, a simplified version of the Control Distance Intersection over Union (CDIoU) loss function accelerates network training. Evaluated on a vineyard obstacle dataset, PRCN achieves a mean average precision (mAP) of 64.3 % and a frame rate of 124.92 frames per second (FPS). Benchmarking against other state-of-the-art models, including YOLOv8n, YOLOv7-Tiny, YOLOv6s, YOLOv3-Tiny, YOLOv4-Tiny, and EfficientDet-D0, demonstrates PRCN's superior performance in balancing accuracy and speed, making it a promising solution for vineyard robot navigation.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序