Real-time robust individual X point localization for stereoscopic tracking

This paper presents a real-time detection and localization method of individual x point features from cluttered background for stereoscopic tracking, using a machine learning approach. Unlike general interest point detectors such as SIFT or SURF, the proposed method is focused on stable and accurate localization of individual specially-marked objects (x points) in complex scenes at frame rate, hence is very suitable for customizing stereoscopic trackers which are widely used in surgical navigation and robotic vision. The x point localization is performed in a cascade manner. First, x point candidates are proposed over the image at cheap cost. Then, a support vector machine is used to classify the candidates according to their image descriptors. Last, a subpixel localization approach is performed to refine the remaining x points followed by a clustering procedure to eliminate duplicated x points. Finally, a stereoscopic tracker using two optical cameras is built to locate x points in the 3D space. Experimental evaluation is performed to show that the proposed method is robust against imaging noise, out-of-plane rotation, and cluttered background. The 2D localization accuracy is evaluated to be a root mean square error of 0.05 pixel with the maximum error of 0.11 pixel, yielding a frame rate of 15 fps with an image size of 1280 × 1024. The 3D localization accuracy by measuring the distance between two x points using the tracker yields a maximum mean error of 0.32 mm.