With the prosperity of artificial intelligence, more and more jobs will be replaced by robots. The future of precision agriculture (PA) will rely on autonomous robots to perform various agricultural operations. Real time kinematic (RTK) assisted global positioning systems (GPS) are able to provide very accurate localization information with a detection error less than $\pm 2$ cm under ideal conditions. Autonomously driving a robotic vehicle within a furrow requires relative localization of the vehicle with respect to the furrow centerline. This relative location acquisition requires both the coordinates of the vehicle as well as all the stalks of the crop rows on both sides of the furrow. This extensive number of coordinate acquisitions of all the crop stalks demand onerous geographical survey of entire fields in advance. Additionally, real-time RTK-GPS localization of moving vehicles may suffer from satellite occlusion. Hence, the above-mentioned $\pm 2$ cm accuracy is often significantly compromised in practice. Against this background, we propose sets of computer vision algorithms to coordinate with a low-cost camera (50 US dollars), and a LiDAR sensor (1500 US dollars) to detect the relative location of the vehicle in the furrow during early, and late growth season respectively. Our solution package is superior than most current computer vision algorithms used for PA, thanks to its improved features, such as a machine-learning enabled dynamic crop recognition threshold, which adaptively adjusts its value according to the environmental changes like ambient light, and crop size. Our in-field tests prove that our proposed algorithms approach the accuracy of an ideal RTK-GPS on cross-track detection, and exceed the ideal RTK-GPS on heading detection. Moreover, our solution package neither relies on satellite communication nor advance geographical surveys. Therefore, our low-complexity, and low-cost solution package is a promising localization strategy as it is able to provide the same level of accuracy as an ideal RTK-GPS, yet more consistently, and more reliably, as it requires no external conditions or hassle of the work demanded by RTK-GPS.

中文翻译：

---

精确农业中自主漫游车的高精度自适应低成本位置传感子系统

随着人工智能的繁荣，越来越多的工作将被机器人取代。精确农业（PA）的未来将依靠自动机器人来执行各种农业操作。实时运动（RTK）辅助的全球定位系统（GPS）能够提供非常准确的定位信息，且检测误差小于$ \ pm 2 $厘米在理想条件下。在沟槽内自主驾驶机器人车辆需要相对于沟槽中心线对车辆进行相对定位。这种相对位置的获取需要车辆的坐标以及犁沟两侧的农作物行的所有茎。所有农作物秸秆的大量坐标采集需要事先对整个田地进行繁重的地理调查。此外，移动车辆的实时RTK-GPS定位可能会受到卫星遮挡的影响。因此，上述$ \ pm 2 $在实践中，cm精度通常会大大降低。在此背景下，我们提出了一套计算机视觉算法，可以与一台低成本相机（50美元）和一个LiDAR传感器（1500美元）进行协调，以检测早期和晚期车辆在犁沟中的相对位置。生长季节。我们的解决方案包优于PA所使用的大多数当前计算机视觉算法，这归功于其改进的功能，例如支持机器学习的动态作物识别阈值，该阈值可根据环境变化（例如环境光和作物大小）自适应地调整其值。我们的现场测试证明，我们提出的算法在跨轨检测中可以达到理想的RTK-GPS的精度，而在航向检测中可以超过理想的RTK-GPS。此外，我们的解决方案包既不依赖卫星通信，也不依靠先进的地理调查。因此，我们的低复杂度，低成本解决方案包是一种有前途的本地化策略，因为它能够提供与理想的RTK-GPS相同的精度，而且更加一致，更可靠，因为它不需要外部条件或RTK-GPS要求的工作麻烦。