A Robotic System Employing Deep Learning for Visual Recognition and Detection of Weeds in Grasslands

Development of User-Integrated Semi-Autonomous Lawn Mowing Systems: A Systems Engineering Perspective and Proposed Architecture

“…A method to detect trash within a grassy environment was proposed by Bai et al [19], using a ground segmentation method and a deep neural network to detect the foreign items. Similarly, a field robotic system to detect a specific species of weeds, also using neural networks to examine 2-D images of the field, was developed by Kounalakis et al [20] Automated and intelligent path planning, as long as it can be shown to be reliable for the needed conditions, can help to find the best coverage for a particular yard or field. These methods may be used to adapt in real-time if the map needs to be updated after beginning the job (such as when the system encounters an unexpected obstacle).…”

Deep learning-based visual recognition of rumex for robotic precision farming

“…Therefore, a lot of attention has been given by the scientific community and several weed control robotic systems have been presented using Unmanned Aerial Vehicles (UAV) [1,2] and Unmanned Ground Vehicles (UGV) [3,4,5]. Such systems rely on accurate and reliable recognition of weeds from visual or other types of sensory data [1,3,6,4,7,8,9,10,11,12,13]. Such weed control systems are able to detect plants that are not part of the preferred flora and apply treatment to individual plants, e.g.…”

“…The robot integrates visual weed recognition from high-resolution 2D images captured in the uncontrolled environment of grasslands. While the majority of relevant literature focuses on crops, the much more challenging setting of grasslands has started being explored as well [7,8,9,10,11,12,13]. The weed to be recognized by our proposed system is the Broad-leaved dock (Rumex obtusifolius L.), a weed with wide geographic distribution that constitutes a major problem for dairy farms in many parts of the world.…”

“…Therefore, several design considerations contributed in the final pipeline of the proposed algorithm. Each considered frame acquired from a top-view camera mounted on the robot [13], is partitioned in several blocks, corresponding to spatial regions, using the proposed experimental framework in [10]. In order to perform the recognition process and compensate for limitation of deep learning techniques to process high-resolution images, we exploit Transfer Learning, where the content of each image is represented with features computed from a pre-trained Convolutional Neural Network (CNN) [15,16,17,18,19].…”

“…On the other hand, the existence of falsenegative results, i.e, the number of existing weeds misclassified as soil or grass, is not considered as critical and can be compensated by multiple robot passes. The experimental assessment is performed on the publicly available Rumex Ancenis dataset 1 acquired in an open-field in France [13].…”