There is an increasing interest in agricultural robotics and precision farming. In such domains, relevant datasets are often hard to obtain, as dedicated fields need to be maintained and the timing of the data collection is critical. In this paper, we present a large-scale agricultural robot dataset for plant classification as well as localization and mapping that covers the relevant growth stages of plants for robotic intervention and weed control. We used a readily available agricultural field robot to record the dataset on a sugar beet farm near Bonn in Germany over a period of three months in the spring of 2016. On average, we recorded data three times per week, starting at the emergence of the plants and stopping at the state when the field was no longer accessible to the machinery without damaging the crops. The robot carried a four-channel multi-spectral camera and an RGB-D sensor to capture detailed information about the plantation. Multiple lidar and global positioning system sensors as well as wheel encoders provided measurements relevant to localization, navigation, and mapping. All sensors had been calibrated before the data acquisition campaign. In addition to the data recorded by the robot, we provide lidar data of the field recorded using a terrestrial laser scanner. We believe this dataset will help researchers to develop autonomous systems operating in agricultural field environments. The dataset can be downloaded from http://www.ipb.uni-bonn.de/data/sugarbeets2016/.
The growing world population calls for more efficient and sustainable farming technologies. Automating agricultural tasks has great potential to improve farming technologies.A key requirement for full automation is the ability of agricultural vehicles to accurately navigate entire fields without damaging value crops. One important precondition for autonomous navigation is localization, that is, the ability of a vehicle to accurately estimate its pose relative to the crops. A majority of localization approaches detect crop rows to track the heading and lateral offset of the vehicle. This is sufficient to guide the vehicle along crop rows while driving inside the field. However, switching between rows requires a longitudinal pose estimate to determine when to turn at the end of the field.Additionally, at the end of the field sensor data contains less crop row structure and more noise from wild growing vegetation. This can lead to false-positive crop row detections. In this paper, we present a localization approach that goes beyond state-of-the-art crop row following algorithms by providing robust pose estimates not only inside the field but also at the end of the field. The underlying concept of our approach is to estimate the vehicle pose relative to a global navigation satellite system (GNSS)-referenced map of crop rows. This allows us to fuse crop row detections with GNSS signals to obtain a pose estimate with the accuracy comparable to a row following approach in the heading and lateral offset, while at the same time maintaining at least GNSS accuracy along the row.Employing a GNSS-referenced map of crop rows poses several challenges. To relate the detected crop rows to those in the map, we propose a data association strategy that finds correspondences between two sets of lines, that is, crop rows. Furthermore, we improve the GNSS-based longitudinal pose estimate by detecting the end of the field from vegetation data. Additionally, we introduce a novel method to determine false-positive crop row detections to increase the overall robustness in particular in challenging scenarios at the end of the field. Extensive real-world experiments on three different types of crops demonstrate that our localization approach is well suited for fully autonomous navigation in entire fields.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.