基于实例分割的轻量化玉米作物行实时检测方法

Real-time Detection of Lightweight Corn Crop Rows Based on Instance Segmentation

摘要:

田间机器人是当下研究热点，机器视觉导航以成本低廉、效果可靠成为农业机器人全自主田间导航重要发展方向。在机器视觉导航线提取方法中，基于图形学的方法存在适用性差、准确率低、易受噪声干扰等缺点，基于目标检测方法受制于作物重叠，识别难度大，基于语义分割的方法检测速率慢、实时性差、易出现作物行缺失的情况。针对上述问题，以玉米田间除草机器人视觉导航为研究对象，本研究提出应用实例分割算法（Lightweight corn row instance segmentation，LCR-IS）玉米作物行检测方法展开试验工作。模型融合轻量化主干网络，引入了边缘注意力融合模块（C2f edge EMA module，C2f-EEMA）加强特征融合能力，同时利用参数共享机制、重参数化原理结合集成卷积模块（Conv_GN）设计了轻量化实例分割检测头（Rep shared-convolutional segmentation head,RSCS）。利用LCR-IS模型完成作物行提取后，通过最小欧氏距离确定作物行中心位置，最后利用最小二乘法拟合作物行中心线为导航线。试验结果表明，LCR-IS模型在精度方面，F1分数为95.3%、平均交并比为86.1%，在导航精度方面平均角度误差仅有1.17°、距离误差为6.10像素，优于基于YOLO v8-seg、ESNet、VGG-UNet、Moblie-UNet、Mobile-DeeplabV3+的导航线提取方法。LCR-IS模型在边缘设备Jetson Nano上推理速度为18.6 f/s，且无需高精度数据标注即可实现较为准确的导航线提取任务。研究结果能满足玉米田间自动除草机器人的实际需求，在农业机器人实时视觉导航领域具有重要的应用前景。

Abstract:

Field robots are a current research hotspot, and machine vision navigation has emerged as a key development direction for fully autonomous field navigation in agricultural robots due to its low cost and reliable performance. Among machine vision navigation line extraction methods, graphics-based approaches suffer from poor applicability, low accuracy, and susceptibility to noise interference. Object detection-based methods face challenges in crop overlap recognition, while semantic segmentation-based methods exhibit slow detection speeds, poor real-time performance, and frequent crop row omission. To address these challenges, focusing on visual navigation for corn field weeding robots, an experimental approach was proposed by using the lightweight corn row instance segmentation (LCR-IS) algorithm for corn row detection. The model integrated a lightweight backbone network and incorporated the C2f-EEMA (C2f edge EMA module) edge attention fusion module to enhance feature fusion capabilities. Simultaneously, it employed a parameter sharing mechanism and the principle of reparameterization combined with an integrated convolutional module (Conv_GN) to design the lightweight instance segmentation detection head (Rep shared-convolutional segmentation head, RSCS). After extracting crop rows using the LCR-IS model, the center positions were determined via minimum Euclidean distance. Finally, the centerlines were fitted as navigation lines by using least squares. The model achieved an F1 score of 95.3% and an average intersection-over-union (IoU) of 86.1%. The extracted predicted navigation lines exhibited an average angular error of only 1.17° and a distance error of 6.10 pixel compared with actual navigation lines, outperforming navigation line extraction methods based on YOLO v8-seg, ESNet, VGG-UNet, Mobile-UNet, and Mobile-DeeplabV3+. The LCR-IS model achieved an inference speed of 18.6 f/s on the Jetson Nano edge device and enabled relatively accurate navigation line extraction without requiring high-precision data annotation. The research results met the practical needs of automated corn field weeding robots and held significant application potential in the field of real-time visual navigation for agricultural robots.