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

Deployment-Oriented Baseline Evaluation of YOLO26n, YOLOX-S, and RT-DETRv4-S for Broccoli Crown Detection

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

Citation:  2026 ASABE Annual International Meeting  2601206.(doi:10.13031/aim.202601206)
Authors:   Joon Ho Chang, Bo Liu
Keywords:   Agricultural robotics, broccoli detection, computer vision, delayed actuation, edge inference, information technology, precision agriculture, RTDETR, selective harvesting, sensors and control systems, vision transformers, YOLO

Abstract. Selective broccoli harvesting requires a perception system that detects crop heads reliably before the cutting mechanism reaches the plant. This manuscript reports the current detection-only stage of a larger autonomous broccoli harvesting project. At this stage, tracking, coordinate transfer to the robotic harvesting trailer, and deep learning-based harvest-readiness selection have not yet been implemented. Three real-time object detectors, YOLO26n, YOLOX-S, and RT-DETRv4-S, were trained and evaluated on a 5,023-image COCO-format subset of a Broccoli dataset containing 5,761 single-class broccoli crown annotations. Using a common COCO-style evaluator on a held-out test split, YOLO26n achieved the highest AP@0.50:0.95 (0.877) and AP@0.50 (0.986), while RT-DETRv4-S achieved the highest AP@0.75 (0.933) and nearly matched YOLO26n in AP@0.50:0.95 (0.874). For deployment-oriented operating points, confidence thresholds were selected on the validation split and then frozen for held-out test evaluation. At these validation-selected thresholds, RT-DETRv4-S achieved the strongest annotated-test operating-point performance, with precision 0.9929, recall 0.9745, F1 0.9836, and 0.0080 FP/frame. However, the soil/leaves-only negative-frame stress test revealed a different deployment tradeoff: YOLO26n produced zero false positives on 648 evaluated negative frames, whereas YOLOX-S and RT-DETRv4-S produced 76 and 273 false positives, respectively. YOLO26n also provided the lowest mean image-space center error (2.42 px) and the best 5 px and 10 px within-threshold localization rates. These results show that detector selection for robotic broccoli harvesting must consider annotated-frame detection accuracy, false-trigger behavior on negative scenes, and center localization rather than relying on a single score.

(Download PDF)    (Export to EndNotes)