6D pose estimation of medicine bottles via RGB-D fusion and SC-YOLOv8 instance segmentation for service robotics

Abstract Accurate pose estimation of medicine bottles is essential for service robots performing healthcare tasks such as automated picking and opening operations. While numerous 6D pose estimation methods have been developed for various objects, existing approaches face challenges when applied to medicine bottles due to their limited distinguishing features, small size, and axisymmetric geometry. This study presents a 6D RGB-D fusion pose estimation method based on the proposed SC-YOLOv8 instance segmentation model. The proposed model improves segmentation performance for small-scale objects with textureless feature through three integrated components: Spatial Depth Convolution Module (SPD-Conv), Convolutional Block Attention Module (CBAM), and an optimized segmentation layer. Instance segmentation experiments conducted on a self-built medicine bottle dataset achieved mAP50 and mAP50:95 scores of 91.30% and 74.28%, respectively, outperforming other instance segmentation models. Additionally, this study employs optimized principal component analysis (PCA) to address axisymmetric pose ambiguities in bottles and caps. The experimental results show that the accuracy rates of pose estimation for the bottles and the caps reach 100% and 50% respectively, and the frame rate of real-time pose estimation reaches 11 FPS, thereby validating the effectiveness and practical applicability of the proposed 6D pose estimation approach. The proposed method provides an effective and practical solution for 6D pose estimation of medicine bottles and supports automated dispensing applications in medical service robotics.