AppleYOLO: Apple yield estimation method using improved YOLOv8 based on Deep OC-SORT

Abstract

The accuracy of apple yield estimates is critical to orchard management. Existing apple yield estimation methods still lack in accuracy and efficiency. To solve this challenge, an Apple yield estimate method based on YOLOv8 and Deep OC-SORT (AppleYOLO) is proposed in this paper. To adequately learn the edge information of the apples, the lightweight FasterNet is used as the backbone portion of the AppleYOLO. In order to make AppleYOLO accurately capture the contextual information and improve the spatial perception of the apples, the Focal Modulation is designed after its backbone. To address the ability to extract complex features, the dynamic convolutional KernelWarehouse with efficient parameters is employed in the feature fusion part of AppleYOLO. For overcoming the problem of unstable tracking and repeated counting when detecting apples in real time, the Deep OC-SORT is deployed in AppleYOLO. In our customized dataset, validated by ablation experiments, the introduction of FasterNet, Focal Modulation, and KernelWarehouse enhances the detection performance of AppleYOLO. In comparison with the benchmark model YOLOv8, AppleYOLO achieves mAP50 and mAP50-95 of 98.5% and 79.8%, respectively, with 1% and 5.1% improvement. Moreover, the performance of AppleYOLO is superior compared to other state-of-the-art methods, such as YOLOv9-t, RT-DETR, YOLOv8 and YOLOv7. The experiments prove that AppleYOLO to be achieve the goal of high accurate and high efficient.