3D Display And Stereo Matching Algorithm Based On Binocular Medical Endoscope System

With the development of stereo vision, acquirement for3D display has become the hotspot in the field of clinic application. When binocular camera system catches two target images with different viewing angles, three-dimensional scene and its depth would be presented to doctors, which can reduce complications after operation. The main work of this thesis is to design software system of our stereo endoscope with calibration, epipolar rectification of binocular images, stereo matching algorithm for depth map.Firstly, we introduced the requirements and performance index of medical endoscope, and built the plan for our3D endoscope system. Then the design of software was presented in detail, including synchronous operation of binocular images capture and display, system parameters calibration and image correction, efficient dense disparity map generation method and related control module.Binocular cameras always have relatively rotating and relative translation as they are not installed exactly parallel. So we researched on parameters calibration and stereo images correction method to realize optimal3D display and rectifying images for stereo matching, which focused on analyzing transformation relationship among each coordinate of binocular system, giving a method to computing world coordinate of target point in real scene, and obtaining stereo camera parameters to finish epipolar rectification module.Finally we discussed local matching methods for stereo matching, including three important existing ideas—color similarity based method, census transformation method, adaptive weighted method. By researching on their matching principle and performance, we proposed improved idea such as edge constraint, simplified Hamming distance computation, direct comparison of weight window respectively, and found that improved adaptive weight algorithm perform best in the situation of both consistent and inconsistent illumination. However, the captured binocular images of endoscope always have different illumination as its optical system has very small diaphragm. This generated different pixel values between corresponding points, which resulted in failure of traditional color model. So we introduced a color normalization representation and proposed an improved matching plan. This new color model can remove component caused by referred illumination factors and gamma correction. We also proposed a grayscale distribution based area segment method to decrease target points. Finally, improved adaptive weighted method was used to compute similarity and WTA was used to search optimum matching points. Experimental results showed our improved method performed better than traditional ones especially in our3D medical endoscope.