Perception Stack for an Agrobot

Robotics for agriculture has seen an increasing amount of interest, particularly in automating tasks that are labour intensive. In this work, we focus on building a vision system for automatically constructing a deformable model of a plant from a given front-facing plant image. We use a learning-based classification step to pre-segment the observed image into stems and leaves and apply a gradient descent based method for fitting a deformable model of the stems onto the observed image. Individual stem portions are stitched together using a directed search to make a complete model. The proposed methodology will serve as a framework for automatic image understanding for agricultural robots in greenhouses