Graph Neural Networks for Plant Organ Tracking

摘要

<b><sc>Abstract.</sc></b> <b>Much progress has been made over the past decade on the problem of multi-object tracking. Many recent techniques leverage Convolutional Neural Networks (CNNs) and are focused on the domain of autonomous driving or people-tracking. In contrast, we concern ourselves with how these recent advances can be adapted to the domain of High-Throughput Phenotyping (HTP). HTP leverages automated sensing capabilities in order to speed up the process of measuring plant phenotypic traits to advance breeding programs. Many specific problems within the domain of HTP require the accurate localization of plant organs, as well as the tracking of the organs over time. Mobile robotic platforms (both ground and air) are typically equipped with localization sensors as well as RGB cameras. We propose a Graph Convolutional Neural Network (GCNN) that is capable of extracting and fusing features from RGB cameras over multiple frames, as well as using these features in a Graph Neural Network to solve the tracking association problem. Our end-to-end tracking approach requires minimal hyperparameters and is easier to train than older approaches that separate affinity computation and track association into two steps. Specifically, we demonstrate our system‘s ability to detect and track individual cotton blossoms in the field, which will be important for both breeding programs and yield estimation.</b>