The feature_tracker package is a versatile tool to detect and track point features in a RGB-D video stream. The framework can be easily modified to use different detection & tracking algorithms. The current implementation uses Kanade-Lucas-Tomasi algorithm (opencv implementation) to first detect corner features on the first frame and then track them in consecutive frames. it maintains a constant number of features by detecting new ones.
Detection and tracking use only RGB data and provide the coordinates in image space. The 3-D Euclidean space coordinates corresponding to each tracked feature is obtained from the registered depth channel.
The feature_tracker can define a ROI between a max and a min allowed depth. It can also use an external mask that we use to define the area of the image occluded by the robot (self-occlussion mask).