03 Mars – Thesis defense - Louis Lac

10 h Amphi Jean-Paul Dom - laboratory IMS (University of Bordeaux | Talence campus)

Computer vision and deep learning methods for locating, tracking and analyzing the structure of plants. Application to precision hoeing.

Agriculture faces today the challenge of steadily increasing food demands, caused by the sustained population growth. At the same time, the environmental crisis, for which agriculture is partly responsible, risks having a negative impact on human societies. Limiting these impacts while ensuring global food security is therefore a necessity. Precision agriculture, which aims to treat crops in a more precise and more localized way than traditional methods of intensive agriculture, is one of the levers of the transition to a more resilient and less polluting agriculture. Recent advances in computer vision, robotics and artificial intelligence make it possible to imagine new innovative solutions to automate certain tasks such as precision hoeing.
In this thesis work, computer vision and deep learning methods are explored in order to detect and collect properties on plants of interest as well as to track crops in sequences of images. The goal is to provide powerful and robust algorithms which can be used in precision hoeing and crop monitoring applications. The proposed work is based on a prototype of a mechanical precision hoeing tool named BIPBIP to experimentally evaluate the developed methods.
Several deep learning methods designed for the localization and characterization of plants are explored, parameterized and evaluated, and an algorithm for tracking plants of interest is evaluated. The experiments presented in this work show that it is possible, on the one hand, to leverage deep neural networks to detect the structure of crops and to precisely locate their organs and, on the other hand, to improve the performance of the detection thanks to a temporal tracking algorithm.

Event localization