文献类型： 外文期刊

作者： Zhang, Jiwei ¹ ; Jiang, Liguo ¹ ; He, Leilei ¹ ; Wu, Zhenchao ¹ ; Li, Rui ¹ ; Chen, Jinyong ⁶ ; Sun, Xiaoxu ⁶ ; Xue, Yunfei ⁹ ; Grecheneva, Anastasia ¹⁰ ; Fountas, Spyros ¹¹ ; Fu, Longsheng ¹ ;

作者机构： 1.Northwest A&F Univ, Coll Mech & Elect Engn, Yangling 712100, Shaanxi, Peoples R China

2.Minist Agr & Rural Affairs, Key Lab Agr Internet Things, Yangling 712100, Shaanxi, Peoples R China

3.Shaanxi Key Lab Agr Informat Percept & Intelligent, Yangling 712100, Shaanxi, Peoples R China

4.Northwest A&F Univ, Shenzhen Res Inst, Shenzhen 518000, Guangdong, Peoples R China

5.Anhui Agr Univ, Sch Engn, Hefei 230061, Anhui, Peoples R China

6.Chinese Acad Agr Sci, Zhengzhou Fruit Res Inst, Zhengzhou 450009, Henan, Peoples R China

7.Natl Key Lab Germplasm Innovat & Utilizat Hort Cro, Zhengzhou 450009, Henan, Peoples R China

8.Chinese Acad Agr Sci, Zhongyuan Res Ctr, Xinxiang 453500, Henan, Peoples R China

9.Meixian Agr Machinery Technol Extens Serv Ctr, Baoji 722300, Shaanxi, Peoples R China

10.Russian State Agrarian Univ, Moscow Timiryazev Agr Acad, Moscow 127434, Russia

11.Agr Univ Athens, Athens 11855, Greece

关键词： Yield estimation; Fruit detection; Regional detection; Multi-object tracking; Fruit counting

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 237 卷

页码：

收录情况： SCI

摘要： Automated kiwifruit counting in orchards delivers accurate, timely, and cost-effective insights into yield estimation, which is crucial for decision-making in harvesting, storage, and marketing operations. Although several studies have proposed methods for kiwifruit counting in orchard, these methods have mostly focused on a small section within the orchard row, which may provide insufficient data for practical application due to the complex and uneven fruit-growing condition. This study introduces a novel automatic approach for row-based accurate kiwifruit counting on video sequences. The sequences are captured by smartphones mounted on a stabilizer and an extension pole, from an upward perspective that spans the full length of each kiwifruit row, addressing the limitations of previous methods. The pipeline comprises of three key components: kiwifruit and support-post detection, video-based kiwifruit counting, and detection region adaptation. First, the performance of You Only Look Once (YOLO) detection models was compared based on the constructed dataset, indicating that the medium-scale model achieved a balanced performance in terms of parameters, model size, inference time and average precision (AP). Second, a two-containers verification (TCV) method was proposed and applied following fruit tracking to reduce over-estimation in video-based counting. Finally, the kiwifruits in neighboring rows were eliminated by detection region adaptation, which estimated the row boundaries and dynamically adapting the masks based on support-posts in orchards. YOLOv5m and YOLOv8m were considered as the most competitive frameworks in kiwifruit detection, achieving AP0.5:0.95 scores of 0.864 and 0.878, respectively. The proposed TCV method suppressed false counts and counting accuracy improved by 59.06 % (from 40.12 % to 94.20 %) on ByteTrack and 54.08 % (from 37.59 % to 96.65 %) on DeepSORT. Moreover, the detection region adaptation guided by support-post has eliminated most fruit counts in neighboring rows. The R-squared (R2) of the row-based kiwifruit counting was 0.9791, indicating the proposed approach has the potential to achieve yield estimation for kiwifruit orchards.