This presentation showcases the feasibility of using 2D video as a state estimator for feedback in the CTR control loop. Evaluated feature types include SURF, SIFT, ORB, KAZE, and BRISK. The objective is to extract affine transformations from frame-to-frame motion in an orthogonal Cartesian system (x, y, z) and rotation around the z-axis along the camera midline. The accuracy of the camera movement algorithm is validated against the NDI Aurora 6 DOF Reference tracker.
The most accurate point feature for the developed tracking algorithm is determined to be the combination of KAZE and SURF type features. Comparing this performance to the accuracy of forward kinematic approaches, with an RMSE of 1.4 mm, it was concluded that using these point feature types can improve pose accuracy based on forward kinematics using image-based feedback.
Following this validation, the ability to extract features from real-world bladder footage and construct affine transformations thereof has been tested. The videos are selected to cover a wide range of scenarios found in the bladder. When tested on real-world videos of a bladder inspection, a combination of the SIFT and KAZE feature types results in the most complete and consistent tracking across different scenarios. The lack of arteries and poor visibility of the bladder wall in one of the scenarios proved to be the most challenging environment for visual odometry using point features.