Abstract
For robotic technology to be adopted within the
agricultural domain, there is a need for low-cost systems that can be deployed autonomously across a wide variety of crop types, environmental conditions, and planting methods, without extensive re-engineering. We present an end-to-end learning approach for row following in agriculture, that can be used for navigation on lightweight robotic platforms. Building on recent work on deep convolutional neural networks (DCNNs) and end-to-end learning approaches, we propose to train our DCNN to output control commands directly from RGB image input data, using a large-scale forest trail dataset and then fine-tune on small datasets from agricultural settings. For this purpose, we recorded data for row-following from a strawberry polytunnel and a sugar cane field. Preliminary evaluation on independent test datasets show promising results on a domain not seen during training. This indicates that our approach generalises well across agricultural domains, and that the low-level features obtained from the trail dataset are relevant for agricultural applications. Future work includes data capture from different applications and seasons to train and test on more data, and verify the control approach on a real robot or drone.
agricultural domain, there is a need for low-cost systems that can be deployed autonomously across a wide variety of crop types, environmental conditions, and planting methods, without extensive re-engineering. We present an end-to-end learning approach for row following in agriculture, that can be used for navigation on lightweight robotic platforms. Building on recent work on deep convolutional neural networks (DCNNs) and end-to-end learning approaches, we propose to train our DCNN to output control commands directly from RGB image input data, using a large-scale forest trail dataset and then fine-tune on small datasets from agricultural settings. For this purpose, we recorded data for row-following from a strawberry polytunnel and a sugar cane field. Preliminary evaluation on independent test datasets show promising results on a domain not seen during training. This indicates that our approach generalises well across agricultural domains, and that the low-level features obtained from the trail dataset are relevant for agricultural applications. Future work includes data capture from different applications and seasons to train and test on more data, and verify the control approach on a real robot or drone.