基于OrchardYOLOP的火龙果园多任务视觉感知方法

Multi-task Visual Perception Method in Dragon Orchards Based on OrchardYOLOP

摘要:

现代果园机器人面临复杂环境、光线多变和非结构化环境等问题，需要高效处理大量环境信息，而传统顺序执行多个单一任务的算法受到计算能力的限制，难以满足现代果园机器人的需求。本文针对火龙果园环境中自动驾驶机器人处理多任务时所面临的实时性和准确性要求，基于YOLOP模型引入了焦点融合高效卷积模块，并采用C2F和SPPF模块，同时优化了分割任务的损失函数，从而构建出OrchardYOLOP模型。实验结果表明，在目标检测任务上的精确度达到84.1%；在可行驶区域分割任务上的mIoU达到89.7%；在果树区域分割任务上的mIoU提高到90.8%；推理速度达到33.33f/s，而参数量仅有9.67×106。与YOLOP模型相比，不仅在速度上满足了实时性要求，而且准确性上也有显著提升。这解决了火龙果园多任务视觉感知中的关键问题，为非结构化环境下的多任务自动驾驶视觉感知提供了一种有效的解决方案。

Abstract:

In the face of challenges such as complex terrains, fluctuating lighting, and unstructured environments, modern orchard robots require the efficient processing of a vast array of environmental information. Traditional algorithms that sequentially execute multiple single tasks are limited by computational power which are unable to meet these demands. Aiming to address the requirements for realtime performance and accuracy in multitasking autonomous driving robots within dragon fruit orchard environments. Building upon the YOLOP, focus attention convolution module was introduced, C2F and SPPF modules were employed, and the loss function for segmentation tasks was optimized, culminating in the OrchardYOLOP. Experiments demonstrated that OrchardYOLOP achieved a precision of 84.1% in target detection tasks, an mIoU of 89.7% in drivable area segmentation tasks, and an mIoU increased to 90.8% in fruit tree region segmentation tasks, with an inference speed of 33.33 frames per second and a parameter count of only 9.67×106. Compared with the YOLOP algorithm, not only did it meet the real-time requirements in terms of speed, but also it significantly improved accuracy, addressing key issues in multi-task visual perception in dragon fruit orchards and providing an effective solution for multi-task autonomous driving visual perception in unstructured environments.