A Novel Reconstruction Model of Multi-camera Positioning System Based on Neural Network

Abstract

Optical positioning based on multiple cameras could obtain superior performance with large measuring volume and excellent accuracy, and is therefore widely employed in medical robotics, especially in surgical robotics. The optical positioning precision highly depends on the three-dimensional (3D) reconstruction model, and the traditional Euclidean reconstruction is tanglesome in calibration and is time-consuming for a real-time system. This paper proposed a novel Radial Basis Function Neural Network (RBFNN)-based reconstruction model without complicated calibration and with the capacity of approximating nonlinear functions for multi-camera positioning system (MCPS). Moreover, a special image pre-processing was also proposed for improving the stability and precision of RBFNN-based reconstruction model. The experimental results revealed the proposed reconstruction model could achieve superior performance with accuracy of 0.2676mm and is robust in both static and dynamic scenarios.