Keypoint-based Framework for Multi-instance Instrument Pose Estimation in AR Surgical Navigation

In surgical navigation systems, accurately identifying and localizing the spatial positions of surgical instruments is the foundation of scene perception and human-computer interaction. However, marker-based methods and external devices face limitations in achieving precise instrument tip pose due to sterilization requirements and operating range constraints. To address the challenge of vision-based instrument pose estimation in endoscopic scenarios, we propose a novel pose estimation framework with a keypoint-based network and the Perspective-n-Point method. Experimental results on the SurgRIPE dataset show that our proposed keypoint generation strategy and dataset synthesis technique can greatly improve the performance of our framework, which finally achieves an outstanding result with a real-time speed of 5.5 ms.