Motion Vector for Outlier Elimination in Feature Matching and Its Application in SLAM Based Laparoscopic Tracking

摘要

This paper presents a motion vector-based method to detect and remove the outlier of the matched feature point in laparoscopic images. Feature point detected on organ surface in laparoscopic images plays an important role not only in laparoscopic tracking but also in organ surface shape reconstruction. However, many factors such as the deformation of the organ or the movement of the surgical tools result to the outliers in matched feature points, thus the feature point based tracking and reconstruction will have larger errors. Traditional methods use these points either directly (inside a RANSAC scheme) or after a prior knowledge of compensation, which may lead to larger error in tracking and reconstruction. We introduce the motion vector (MV) based method to detect outliers among the matched feature points. MV is originally used in the compression of the video streams, we exploit it to detect the movement of one feature point in different video frames. The outliers of feature point can be detected by enforcing a direction constraint with its MV. Our method had been implement under a SLAM-based framework for laparoscopic tracking, we modified the map management of SLAM for better laparoscopic tracking. The experimental results showed that our method effectively detects and removes the outliers without any prior knowledge; the average precision rate in image pairs was 95.9%.