Detecting tomato crops in greenhouses using a vision based method

Precision Farming is generally defined as an information and technology basedfarm management system that allow to identify, analyse and manage variability within fields for optimum profitability, sustainability and protection of the land resource (Singh). The use of intelligent systems could significantly contribute to increase overall performances in intensive culture management and production efficiency, reducing costs, and, not least, to improve labour quality and safety. This work targets the problem of the implementation of a machine vision algorithm that allow tomato detection in images and, as future development, once it will be integrated with arobot (Balloniet al. 2009, Longo et al. 2010, Schillaciet al. 2009), it will be really useful to perform precision farming activities: fruit classification, harvesting, local chemicals treatment etc. The overall tomatoes detection architecture is built around a method for classifying individual imageregions. This is divided into two phases. The off-line learning phase creates a binary classifier thatprovides object/non-object decisions for fixed sized image regions (“windows”) while the on-linedetectionphase uses the classifier to perform a dense multi-scale scan reporting preliminary objectdecisions at each location of the test image. These preliminary decisions are then fused to obtain the final object detections. The approach is data-driven and purely bottomup using low-level visual features to detect objects. The performance of the method depends strongly on the dataset creation (as much bigger is the dataset as better the machine vision will know about a tomato) and detector choice (which is the best representation of the tomato class?).