ORIENTATION OF POINT CLOUDS FOR COMPLEX SURFACES IN MEDICAL SURGERY USING TRINOCULAR VISUAL ODOMETRY AND STEREO ORB-SLAM2
In photogrammetry, computer vision and robotics, visual odometry (VO) and SLAM algorithms are well-known methods to estimate camera poses from image sequences. When dealing with unknown scenes there is often no reference data available and also the scene needs to be reconstructed for further analysis. In this contribution a trinocular visual odometry approach is implemented and compared to stereo VO and ORB-SLAM2 in an experimental setup imitating the scene of a knee replacement surgery. Two datasets are analysed. While a test-field provides excellent conditions for feature detection algorithms with its artificial texture assembled, extracted images show the knee joint itself solely in order to use only the homogenous, but in real application stable, region of the knee joint. The camera trajectories of VO and ORB-SLAM2 are transformed to corresponding coordinate systems and are subsequently evaluated. The tracking algorithms show poor quality when only the inappropriate surface of the knee is used but perform well when the artificial texture of the test-field is used. The third camera does not lead to a significant advantage in this setup using our implementation. Possible reasons, e.g. less overlap, are discussed in this contribution. Nevertheless, the quality of the oriented point clouds, obtained by trinocular dense matching, is less than 1mm for most of the analysed data. The experiment will be used to focus on further developments, e.g. dealing with specular reflections, and for evaluation purposes using different SLAM/ VO algorithms.