Pose Detection and Localization of Pineapple Fruit Picking Based on Improved Litehrnet

Pinlan ChenBin YanGanran Deng^*Guojie Li^*Zhende CuiShuang ZhengFengguang HeLing LiXilin WangSili ZhouShuangmei QinZehua LiuYe Dai

• Agricultural Machinery Research Institute, Chinese Academy of Tropical Agricultural Sciences, Zhanjiang, China

To achieve accurate detection of the pineapple fruit picking area and gesture in complex background environments and under varying lighting conditions, a pineapple keypoints detection model, LTHRNet, based on an enhanced version of LiteHRNet, was developed in this study. Pineapple fruit image data were collected under different lighting conditions, and six keypoints were defined to characterize the morphological features of the pineapple fruit. In the model design, the LKA_Stem, D-Mixer, and MS-FFN modules were incorporated into the LiteHRNet network structure to enhance feature extraction, capture both global and local features, and enable multi-scale feature fusion, respectively. The enhanced LTHRNet model retains high-resolution feature extraction and improves both the precision and spatial accuracy of keypoint detection by integrating subnetworks with varying resolutions in parallel. The experimental results demonstrated that LTHRNet performed well in pineapple keypoint detection tasks. For the KAP0.5 and KAR0.5 indices, LTHRNet achieved 93.5% and 95.1%, respectively. The model exhibited superior detection accuracy and robustness compared to others under complex lighting and occlusion conditions, with a detection speed of 21.1 fps. In pose estimation, the average deviation angle of LTHRNet was 2.37°, significantly lower than that of other models. These results indicate that the LTHRNet model proposed in this study excels in both keypoint detection and pose estimation for pineapples, providing precise keypoint localization and pose estimation data for pineapple harvesting, and offering valuable technical support for the pose localization of other fruits.