HP-YOLO: A Lightweight Real-Time Human Pose Estimation Method

Abstract

Human Pose Estimation (HPE) plays a critical role in medical applications, particularly within nursing robotics for patient monitoring. Despite its importance, HPE faces several challenges, including high rates of false positives and negatives, stringent real-time requirements, and limited computational resources, especially in complex backgrounds. In response, we introduce the HP-YOLO model, developed using the YOLOv8 framework, to effectively address these issues. We designed an Enhanced Large Separated Kernel Attention (ELSKA) mechanism and integrated it into the backbone network, thereby improving the model’s effective receptive field and feature separation capabilities, which enhances keypoint detection accuracy in challenging environments. Additionally, the Reparameterized Network with Cross-Stage Partial Connections and Efficient Layer Aggregation Network (RepNCSPELAN4) module was incorporated into the detection head, boosting accuracy in detecting small-sized targets through multi-scale convolution and reparameterization techniques while accelerating inference speed. On the COCO dataset, our HP-YOLO model outperformed existing lightweight methods by increasing average precision (AP) by 4.9%, while using 18% fewer parameters and achieving 1.4× higher inference speed. Our method significantly enhances the real-time performance and efficiency of human pose estimation while maintaining high accuracy, offering an optimal solution for applications in complex environments.