ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Weakly-Supervised Multi-Object Tracking for Real-Time Blueberry Fruit Mapping

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2601104.(doi:10.13031/aim.202601104)
Authors:   Daniel J Petti, Changying Li
Keywords:   Multi-object tracking, robotics, high-throughput phenotyping, active learning, weakly-supervised learning, graph neural network

Abstract. Multi-Object Tracking (MOT) of blueberry fruits and flowers is crucial to enhancing breeding efforts but is extremely difficult due to frequent occlusions and intense appearance variations. Though deep learning techniques have significantly advanced the state-of-the-art of MOT in the past few years, such methods have the disadvantage of requiring complicated data annotations in order to train, which are highly challenging to produce for blueberry data. To overcome this, we propose a weakly-supervised data pipeline that can generate usable tracking annotations with only a trained detector. We propose a novel active learning framework for training this detector and demonstrate its efficacy on a difficult blueberry detection problem. Then we leverage the data pipeline to create a large and diverse dataset which we use to train a tracking model that is effective on a challenging blueberry tracking benchmark. This is an important step towards creating a tracking foundation model that can be deployed in novel situations without re-training. The proposed tracking framework outperforms model-free baselines on a challenging blueberry tracking benchmark. At the same time, it can be incorporated as part of a 2-stage tracking method that remains lightweight enough to maintain real-time performance on edge devices. This 2-stage approach improves on the performance of a single stage, reducing ID switches by 20%, while increasing HOTA and IDF1 by 7 and 15%, respectively. In summary, this study proposes a new paradigm for agricultural MOT that minimizes data requirements and is designed from the ground-up for deployment on field robots.

(Download PDF)    (Export to EndNotes)