2020 年 38 巻 3 号 p. 116-125
In this paper, a local feature extraction method based on fast image guided filtering for organ tracking in Water-filled Laparoendoscopic Surgery (WaFLES) is presented. Robust feature extraction and matching is important for feature based tracking and 3d-reconstruction techniques. We created a non-linear scale space and extracted Hessian determinant features, described with MLDB-descriptor that performed robustly in WaFLES images, outperforming Akaze in matching. We showed that non-linear scale space methods, especially our method, are most robust for WaFLES, over de facto SIFT, SURF and ORB methods. It was demonstrated that the scale space could be created more efficiently using fast guided filtering over non-explicit diffusion. We are bringing these techniques closer to real-time capability in simpler hardware, without sacrificing performance.