Robotics for weed control: I-Weed Robot for a specific spraying

International audienceTo preserve environment for a sustainable agriculture, we explore the development of a new autonomous robot, called I-Weed Robot (Intelligent Weed Robot), which aims at reducing herbicides in crop fields (maize, sunflower...). Using a high precision positioning signal (RTK) to locate the robot in the field, a Kaman filter and a proportional-integral-derivative controller (PID controller) allow adjusting the orientation of the robot depending on a predefined trajectory. As for the spraying system, a camera in front of the mobile platform detects weed plants thanks to an image processing based on a crop/weed discrimination algorithm (Hough Transform). At the back a spray boom triggers the right nozzle at the right time depending on the location of weeds. This article assesses the performance of the guidance and weed detection algorithms using numerical simulations (virtual trajectory, virtual field image). The robustness of the guidance algorithm is tested for different noisy signals (GPS, DGPS and RTK). The accuracy of the crop/weed discrimination algorithm is evaluated using a large data base of synthetic images generated by the 'SimAField' software. To evaluate the accuracy of the algorithm and understand the sources of misclassification errors, the results are summarized in a confusion matrix which indicates the number of correctly and incorrectly classified pixels (both weed and crop classes). Results indicate accuracy up to 90%. A significant number of weed pixels are always considered as crops so that crop detection is overestimated. The reason is that the bandwidth value of the crop row mask that is automatically generated overlaps weed plant close to the crop row. These simulations demonstrate that both algorithms are reliable. However, further research in field conditions is necessary to confirm the promising results