文献类型：期刊文献

英文题名：OALDNet: An orientation-aware lightweight detection network for precision pineapple harvesting

作者：Yang, Ziming[1,2];Shan, Zhe[3];Hao, Xiaying[2,4];Lin, Cong[1];Xue, Zhong[2,4]

机构：[1]Guangdong Ocean Univ, Sch Elect & Informat Engn, Zhanjiang 524088, Peoples R China;[2]South Subtrop Crops Res Inst, Chinese Acad Trop Agr Sci, Key Lab Trop Fruit Biol, Minist Agr, Zhanjiang 524091, Guangdong, Peoples R China;[3]Hainan Univ, Sch Comp Sci & Technol, Haikou 570228, Peoples R China;[4]Shanxi Agr Univ, Coll Agr Engn, Taigu 030801, Shanxi, Peoples R China

年份：2026

卷号：200

外文期刊名：APPLIED SOFT COMPUTING

收录：SCI-EXPANDED(收录号：WOS:001771849900001)、、EI(收录号：20262020718575)、WOS

基金：This work was funded by the Chinese Academy of Tropical Agricultural Sciences for Science and Technology Innovation Team of National Tropical Agricultural Science Center (No. CATASCXTD202513) , Hainan Province Science and Technology Special Fund (No. ZDYF2023XDNY058, No. ZDYF2025XDNY099) , supported by Hainan Provincial Natural Science Foundation of China (325QN432) , Top Ten Guangdong Province Agricultural Science and Technology Innovation Main Attack Directions "Unveiling and Leading" Project (No. 2022SDZG03) , Central Public-interest Scientific Institution Basal Research Fund (No. 1630062022005, No. 1630062025018) , and in part by the Undergraduate Innovation Team Project of Guangdong Ocean University under Grant CXTD2024011 and Grant JDTD2024003.

语种：英文

外文关键词：Agricultural automation; Oriented object detection; Pineapple posture perception; Lightweight network; Edge devices

外文摘要：Mechanized harvesting is fundamental to reducing labor costs and improving efficiency in pineapple production. Accurate orientation estimation is essential for robotic arms to achieve precise grasping, making oriented object detection a key technology for agricultural automation. However, current automated pineapple harvesting sys tems rely on horizontal vision detection, which hinders accurate localization of grasping and cutting points and increases the risk of fruit damage. Moreover, existing oriented detection algorithms are often computationally intensive and structurally complex, limiting their practical deployment on resource-constrained edge devices. To address these challenges, we propose OALDNet, an orientation-aware lightweight detection network designed for real-time pineapple localization and orientation estimation in natural field conditions. OALDNet integrates a compact backbone optimized for efficient inference on edge devices with a quantized multi-scale feature fusion module that enhances the recognition of occluded and arbitrarily oriented pineapples while reducing computa tional costs. To fill the gap in training resources, we also develop a large-scale dataset of pineapples annotated with orientation information, covering diverse orchard conditions such as dense foliage, irregular fruit distribu tion, and variable lighting. Extensive quantitative and qualitative evaluations demonstrate that OALDNet achieves a favorable trade-off between detection accuracy and computational efficiency. Specifically, the proposed model attains AP50 = 0.941 and AP0.5-0.95 = 0.745, with an average orientation error of 4.92 degrees. On the Jetson Orin Nano, OALDNet achieves real-time deployment at 113.32 FPS under TensorRT FP16 with 640 & times;640 input, while main taining only 0.8M parameters. Field experiments confirm its robustness and practical applicability, establishing a strong foundation for intelligent robotic harvesting systems and other precision agriculture applications.