Abstract:Addressing the common challenge of real-time, high-precision localization of tiny, occluded, and fast-moving targets in photoelectric imaging systems, conventional pose estimation models often fail to balance the detection of subtle features with lightweight deployment. This paper takes the picking point localization of tender tea shoots in complex tea gardens as a representative scenario, proposing a lightweight pose estimation model based on an improved YOLOv8n-Pose. First, a parallelized patch-aware attention module is designed to enhance the capturing of delicate features at junctions. Second, a multi-source feature fusion module and a progressive bidirectional recurring neck architecture are constructed for efficient full-scale feature interconnection. Finally, a slide-weighted loss function based on IoU is introduced to focus learning on hard samples. Experiments on a custom dataset show a keypoint mAP of 79.1%, with only 2.59M parameters and 8.3 GFLOPs. Deployed on an edge device with TensorRT, it achieves 106.84 FPS, offering a new paradigm for accurate, lightweight, and efficient real-time localization.