A Framework for Calibrating and Benchmarking Computer Vision Algorithms in Surgical Robotics

The use of computer vision techniques in the medical field has been commonplace since the 1960’s when image analysis was used for chromosome pairing [1]. The used of these techniques in recent years has grown to encompass endoscopy, tool inventory recording, and surgical tool tracking in vivo. Given the potentially life altering stakes for these medical applications, a high degree of accuracy is required of any computer vision algorithm. One of the key elements in threedimensional algorithms is accurately converting image values (in pixel space) into real world Cartesian coordinates. This conversion requires an accurate and straightforward calibration routine in order to identify the correct transformation. Once the calibration is complete it is also important for the developers to know the accuracy. While there are existing calibration routines which follow a linear (pinhole) model [2,3], these methods do not work well for varied camera optics and do not handle image distortion well given the linear transformation matrix. Standard stereo correspondence methods were implemented by Dockter and Kowalewski [4] and were found insufficient in terms of speed and accuracy. Given the aforementioned needs, we present a method for calibration and benchmarking for stereo computer vision algorithms which utilizes basic printable objects and a standard machinists lathe and can accommodate more complex camera modeling. Both the calibration and benchmarking boards can be produced using 3D printing techniques and mounted on a machinists lathe thus making high accuracy calibration achievable using available means. This method is directly applicable to the high-distortion stereo optics common in surgical robotics.