Binocular Computer Vision Measurement And Tracking Research Based On The ORB Algorithm

Since the binocular stereo vision has the superiority on the deep informationmeasurement, it becomes a research hotspot in recent years. Image matching forbinocular stereo vision is not only a key step of it, but also is the most time-consumingphase. Therefore, the research of an algorithm with high matching accuracy and lowtime complexity has always been a hot topic among scholars.Firstly, Zhang Zhengyou’s method is studied and selected as the stereo cameracalibration method. After that this paper introduces the epipolar constraint problem instereo matching. In the stage of image feature extraction and image matching, this paperuses a new algorithm ORB(oriented FAST and rotated BRIEF) which was proposed in2011,ICCV. It is superior to the SIFT and SURF, since matching speed is improvedmore than an order of magnitude under the similar matching accuracy. In the matchingoptimization stage, this paper chooses the PROSAC（Progressive Sample Consensus）,which is faster than the RANSAC algorithm in removing the outliers, and thenintegrates ORB and PROSAC. Simulation experiments show that the new algorithm isbetter than the traditional methods in matching accuracy and real-time aspect.This paper implements this feature matching algorithm in the MFC and OpenCVenvironment, what is more, makes a software which has four function:(1) measuringthe matching keypoint-pair’s3D coordinate in the left camera coordinates;(2) a newmonocular tracking algorithm based on Kalman prediction model which improves thespeed up to higher15fps by reducing the search region, and learning object method toenhance the robustness;(3) the3D coordinate measurement using un-matchingkeypoint-pair and the length of object measurement in static state, the relative error ofresults is under3%;(4) Real time tracking and measuring object in binocular visioncondition, this paper adds the epipolar geometry constraint into Kalman predictionmodel for improving the prediction accuracy, and achieves a good measurementprecision in a non-close distance.Eventually, based on the new algorithm, this paper:(1) improves the previousfeature-tracking problem in speed, and makes it meet the real-time requirement;(2)solves the previous limiting problems of stereo tracking and measurement in shape,color or motion state, makes it any target with rich feature.