Random ferns for multiple target tracking in microscopic retina image sequences

Accurate and robust tracking of retina in operating microscope images is critical for an augmented reality assistance system for retinal surgery. Most retinal surgeries such as the peeling are performed using hand-held intraocular light and hence the tool and its shadow have two different motions, independent from the motion of the retina. In this paper, we propose multi-object motion estimation in high definition operating microscopic images by using a parallel network of random ferns, followed by RANSAC in order to achieve a simultaneous and robust tracking of the retina, the tool, and the tool shadow. Thanks to the separate tracking of each object, the number of outliers is dramatically reduced and the extracted motions are more accurate and reliable even in complex scenes which are considerably occluded by the tool and its shadow. The proposed method is evaluated on several challenging sequences in comparison with SIFT tracking, direct visual tracking, and single random ferns tracking of the retina. The experimental results show that the proposed method has a significantly higher success rate in comparison to the other three approaches with the accuracy of 4 pixels in tractable frames which is comparable with the intra- and inter-observer error of manual tracking (3.4 and 8.5 pixels, respectively).