Under-Canopy Navigation for an Agricultural Rover Based on Image Data

This paper presents an Image data-based autonomous navigation system for an under-canopy agricultural mini-rover called TerraSentia. This kind of navigation is a very challenging problem due to the lack of GNSS accuracy. This happens because the crop leaves and stems attenuate the GNSS signal and produce multi-path data. In such a scenario, reactive navigation techniques based on the detection of crop rows using image data have proved to be an efficient alternative to GNSS. However, it also presents some challenges, mainly owing to leaves occlusions under the canopy and dealing with varying weather conditions. Our system addresses these issues by combining different image-based approaches using low-cost hardware. Tests were carried out using multiple robots, in different field conditions, and in different locations. The results show that our system is able to safely navigate without interventions in fields without significant gaps in the crop rows. In addition to this, we see as future steps, not only comparing more recent convolutional neural networks based on processing power needs and accuracy, but also the fusion of these vision-based approaches previously developed by our group in order to obtain the best of both approaches.