Research On Camera Tracking Method

Camera tracking is an important task in computer vision, which has a greatsignificance in many applications. Camera tracking recover the trajectory of thecamera in the scene through the images sequence. Camera tracking first need toextract the feature points from image sequence, and match the feature points, and thenselect several key frames in the image sequence, at last solving key-framecorresponding to the camera parameters based on the matching feature points.Camera tracking feature point extraction and matching module consume a lot of timein the entire camera tracking. The accuracy and execution speed of the cameraparameters solving module are mainly dependent on the establishment of the cameramodel and the key-frame selection strategy.This paper presents a fast feature point extraction and matching method forfeature point extraction and matching module. Feature point extraction processcombined with Laplacian pyramid and non-uniform multi-directional filter bank.Through image filtering and extracting the local extreme points, the method canquickly extract feature points and retain the scale and direction. First construct theLaplacian pyramid of the original image to obtain scale information of the image,which can retaining the orientation of the image information, and then use anon-uniform multi-directional filter bank decomposition of the pyramid image indifferent directions, in the decomposition of the image extraction local extreme pointas a candidate set of feature points, the use of specific consolidation strategycombined candidate feature points the resulting set of feature points, and the directionvector of the distribution of feature points under the direction of the filter bank. Usethe binary descriptor to describe the feature point, and to assign the direction vectorfor the feature point in accordance with the direction of the feature point candidates.The matching characteristic point pair is determined by the minimum distance and thesecond distance, using mismatching is avoided to some extent, while using theRANSAC algorithm to eliminate the false matching characteristic point pair mismatching threshold. Compare with the existing algorithms the experimentdescribed in this article shows the purposed method has improvement in processingspeed and ensure the matching rate and correct rate characteristic point extraction andmatching algorithm.In the camera parameters solving module we build a accurate camera model, andbased on the feature point matching results to select key-frame, then solving thecamera model for key-frame, the purposed method ensure the accuracy of the premiseeffectively improve camera tracking speed. In solving the camera intrinsic parameters,two strategies are applied to solve the camera tracking technology environment.Feature point matching results in the entire video sequence will be sorted by theirtrajectory life-time and stratification, the key-frame extraction by the stratifiedtrajectory. Solving the key-frame corresponding camera parameters not only improvethe accuracy of the solution of the camcorder, but also effectively improve thetracking speed of the entire camera. Experiment shows that the algorithm can quicklyand effectively recover the camera trajectory.