Automatic feature detection and correspondence in a stereo-vision application

This paper describes the detection of artificially planted markers in intraoperative stereo-images. The objective was to detect these markers and match them between stereo-images so that camera calibration could be used to reconstruct their location and orientation in three dimensions. As the markers were rigidly attached to vertebrae, this allowed surgeons to monitor and record corrections made to scoliotic spines during surgery. A model-based system with multiple templates was implemented for feature detection. The three aspects of template matching-similarity measures, template design and decision criteria were investigated. This included an assessment of different cross-correlation measures. Templates, designed using two image properties-gray levels and edges, were organised in a tree structure for each property. Results and accuracy of detection on real operating room images are presented