BIPBIP: a mechanical and automated intra-row weeding solution

In vegetable crops, mechanical weeding is an alternative to the use of chemicals. However, the currently used mechanical techniques increase labour time and are unable to weed into the crop row, except in the case of well-developed and regularly spaced plants such as lettuce or cabbage. Automation and an increase in the precision of mechanical weeding could address these issues. This study presents BIPBIP, an automated in-row weeding module set up at the back of a robot. The module is designed to weed only one row and to be replicated to weed several rows in parallel. It consists of a vision system and mechanical tools operated by artificial intelligence. The vision system comprises an RGB camera and a LiDAR under a vision tunnel with controlled light conditions. A deep neural network adapted to real-time object detection is used to identify and precisely locate crop stems. Then, an aggregation algorithm is applied to the results of consecutive frames to improve the detection and guide the module along the crop row. Crop and stem locations are then sent to a controller that activates a small pair of hoes around the crop stems. Detection algorithms and mechanical weeding were tested during the Challenge ANR ROSE on maize and beans. For the detection task, precision results were between 85 and 91% and recall results between 82 and 85%. For the overall weeding, 88% of the crops were intact and 71% of the weeds present in the row were hoed. Next step for BIPBIP will be agronomic trials on leeks and other crops such as onions, carrots, and garlics. The final goal is to make BIPBIP fully effective on a wide variety of crops and sell it on a large scale.