Enhancing peanut pre‐sizing with PodTracker: A multiple pod tracking algorithm

Abstract In this work, we introduce novel computer vision and machine learning methods—including instance segmentation (Mask R‐CNN [Region‐Based Convolutional Neural Network]), multiple object tracking (PodTracker and DeepSortMask), and classification (Decision Tree)—to improve peanut pod pre‐sizing at grading stations and evaluate the sizing error associated with conventional pre‐sizer mechanical rollers. The trained Mask R‐CNN model achieved high bounding box average precision (AP) scores, with on Farmer Stock Fancy pods and on No. 1 pods. PodTracker (, where RMSE is root mean square error) outperformed DeepSortMask () in counting and tracking pods across frames. The classification model achieved an F1 score across No. 1, Fancy, and Jumbo pods, exhibiting lower standard deviations () compared to the conventional mechanical pre‐sizer (). These results demonstrate that low‐cost scanning solutions, when combined with advanced computer vision techniques, can be effectively integrated into conventional pre‐sizing machinery to offer more accurate sizing distributions than traditional methods.