An Application Driven Comparison of Several Feature Extraction Algorithms in Bronchoscope Tracking During Navigated Bronchoscopy

This paper compares Kanade-Lucas-Tomasi (KLT), speeded up robust feature (SURF), and scale invariant feature transformation (SIFT) features applied to bronchoscope tracking. In our study, we first use KLT, SURF, or SIFT features and epipolar constraints to obtaininterframe translation (up to scale) and orientation displacements and Kalman filtering to recover an estimate for the magnitude of the motion (scale factor determination), and then multiply inter-frame motion parameters onto the previous pose of the bronchoscope camera to achieve the predicted pose, which is used to initialize intensity-based image registration to refine the current pose of the bronchoscope camera. We evaluate the KLT-, SURF-, and SIFT-based bronchoscope camera motion tracking methods on patient datasets. According to experimental results, we may conclude that SIFT features are more robust than KLT and SURF features at predicting the bronchoscope motion, and all methods for predicting the bronchoscope camera motion show a significant performance boost compared to sole intensity-based image registration without an additional position sensor.