Abstract
This paper presents the challenges agricultural robotic harvesters face in detecting and localising fruits under various environmental disturbances. In controlled laboratory settings, both the traditional HSV (Hue Saturation Value) transformation and the YOLOv8 (You Only Look Once) deep learning model were employed. However, only YOLOv8 was utilised in outdoor experiments, as the HSV transformation was not capable of accurately drawing fruit contours. Experiments include ten distinct fruit patterns with six apples and six oranges. A grid structure for homography (perspective) transformation was employed to convert detected midpoints into 3D world coordinates. The experiments evaluated detection and localisation under varying lighting and background disturbances, revealing accurate performance indoors, but significant challenges outdoors. Our results show that indoor experiments using YOLOv8 achieved 100% detection accuracy, while outdoor conditions decreased performance, with an average accuracy of 69.15% for YOLOv8 under direct sunlight. The study demonstrates that real-world applications reveal significant limitations due to changing lighting, background disturbances, and colour and shape variability. These findings underscore the need for further refinement of algorithms and sensors to enhance the robustness of robotic harvesters for agricultural use.
Abstract (translated)
本文介绍了农业机器人收割机在各种环境干扰下检测和定位水果时所面临的挑战。在受控实验室环境中，既使用了传统的HSV（色相饱和度值）转换方法，也采用了YOLOv8（你只需看一次）深度学习模型。然而，在户外实验中仅使用了YOLOv8，因为HSV转换方法无法准确描绘出水果的轮廓。实验包括十种不同的水果图案，其中包含六只苹果和六只橙子。为了将检测到的中心点转化为三维世界坐标，采用了网格结构进行同态（透视）变换。这些实验评估了在不同光照和背景干扰条件下水果的检测与定位情况，结果表明，在室内环境下表现准确，但在户外环境下面临重大挑战。我们的结果显示，在使用YOLOv8的情况下，室内实验实现了100%的检测准确性，而室外条件下的性能有所下降，平均而言，在直接阳光下时YOLOv8的表现为69.15%的精度。这项研究展示了真实世界的应用揭示了由于光照变化、背景干扰和颜色与形状变异所导致的重要限制。这些发现强调了为了增强农业用途中机器人收割机的鲁棒性，进一步完善算法和传感器的需求。
URL
https://arxiv.org/abs/2502.12403
