The positional accuracy of a near-infrared (NIR) dynamic navigator is remarkably affected by two factors. One is the calibration accuracy of the navigator’s two NIR cameras, and the other is the accuracy of feature point extraction. The current lack of accurate calibration devices for NIR cameras limits further application. Therefore, in this study, an NIR camera calibration device was designed by placing the NIR light source and heat dissipation system at the back of a Halcon transparent glass template. Usage of camera calibration in a specific band and the gray centroid method based on elliptic boundary to extract feature points can further improve the accuracy of vision system calibration and measurement. Repeated tests and verifications showed that the reconstruction accuracy (<0.1 pixels) of the binocular vision system calibrated by the NIR calibration device in a specific band was better than that calibrated by traditional methods.