Assessment of the tomato cluster yield estimation algorithms via tracking-by-detection approaches

Abstract

Automated vision-based detection and counting are critical for accurate tomato yield estimation, which contribute to precise yield management strategies and an efficient food supply chains. Special conditions, including background clutter, occlusion, and varying sunlight, affect the accuracy of crop detection and counting. To determine the most suitable algorithms for this yield estimation context, we herein establish a public multi-object tracking (MOT) dataset for tomato cluster counts, while evaluating and comparing state-of-the-art target detection and MOT-based algorithms. The evaluated detectors consist of YOLOv8 and RT-DETR, which represent algorithms that achieve a balance between accuracy and speed. The tracking algorithms included state-of-the-art methodologies such as SORT, DeepSort, ByteTrack, and BotSort. Initially, the performance of the detectors was rigorously evaluated, followed by a comprehensive assessment of the four tracking algorithms within a multi-target tracking database tailored for this research and structured in the MOT context. The findings reveal that YOLOv8 and RT-DETR achieve 93.6% and 94.9% results at mAP@75, respectively, with RT-DETR exhibiting fewer false detections. When combined with the RT-DETR detector, the ByteTrack-based algorithm registers the highest counting accuracy at 95.5%, whereas BotSort achieves the highest MOTA score with 84.6%. Notably, the trackers without the ReID module (i.e., SORT and ByteTrack) demonstrate greater adaptability to frame rate variations in the test videos. At a 30-fps frame rate, the incorporation of ReID modules in DeepSort and BotSort algorithms significantly enhances the MOTA metric. Looking ahead, we plan to leverage these algorithms into an autonomous inspection platform that aims to estimate crop yield in real-time.