LTK-YOLO: Lightweight Keypoint Detection for Robotic Tea Bud Picking

In unstructured tea garden environments, the high color similarity between tea buds and surrounding leaves, together with severe clustering and occlusion, significantly reduces the localization accuracy of picking points and limits the operational efficiency of tea-picking robots. Existing studies often improve detection performance by increasing network depth, but overlook the deployment constraints of resource-limited edge devices. To address these challenges, this study proposes a lightweight end-to-end tea bud picking point detection model based on YOLOv8-Pose, named LTK-YOLO. The proposed model integrates object detection and keypoint localization within a unified framework. First, a WTConv operation is embedded into the bottleneck of the C2f module to construct a C2f-WTConv structure, enabling wavelet-domain feature extraction that enhances contextual representation while reducing redundant computation and memory access. Second, the adaptive upsampling operator DySample replaces conventional interpolation-based upsampling in the feature fusion stage, improving multi-scale feature reconstruction through learnable point sampling. Third, a weighted Shape-NWD loss function is introduced to incorporate bounding box shape and scale information, which accelerates convergence and improves keypoint localization accuracy for small and occluded targets. Experiments were conducted on a self-built tea bud dataset containing 4000 images, divided into training, validation, and test sets at a ratio of 8:1:1. The proposed LTK-YOLO achieved an F1 score of 91.95% and a mAPkp50 of 92.87%, with an average Euclidean localization error below 15 pixels. Compared with the baseline YOLOv8s-Pose, LTK-YOLO improved APkp50 by 3.56 and 4.02 percentage points for kp1-1 and kp1-2, respectively, reduced Params by 53.10%, and GFLOPs by 32.75%. In addition, the model achieved an inference speed of 85.86 FPS on an NVIDIA RTX 4060 GPU, demonstrating strong real-time capability. Because the dataset was captured in a single spring season at two tea gardens in Hangzhou, the reported validation results are bounded to comparable tea varieties, canopy density, and illumination and occlusion conditions; deployment in other climatic regions, seasons, or on unseen tea varieties is expected to require fine-tuning with locally collected data. Overall, LTK-YOLO achieves a favorable balance between detection accuracy, computational efficiency, and lightweight design, providing a promising solution for real-time tea-picking robotic systems.