Real-time camera tracking method for dynamically-changed scene

Abstract

The invention discloses a real-time camera tracking method for a dynamically-changed scene. According to the method, the camera pose can be tracked and solved stably in a scene changed dynamically and continually; at first, feature matching and camera parameter estimation are performed, and then, scene update is performed, and finally, when the method is actually applied, foreground and background mutil-thread coordinated operation is implemented, the foreground thread is used to perform feature point matching and camera motion parameter estimation on each frame, and the background thread is used to perform KD tree, key frame and three-dimensional point cloud maintenance and update continually, and optimize key frame camera motion parameters and three-dimensional point positions in a combined manner. When the scene is changed dynamically, the method of the invention still can be used to perform camera tracking on a real-time basis, the method of the invention is significantly better than existing camera tracking methods in tracking accuracy, stability, operating efficiency, etc.

Description

technical field [0001] The invention relates to a real-time camera tracking method, in particular to a real-time camera tracking method in a dynamically changing scene. Background technique [0002] Camera tracking technology is one of the important technologies in the field of computer vision, and it has extremely important applications in the fields of robot navigation and augmented reality. [0003] The difficulty of camera tracking technology lies in: 1) Most methods track feature points frame by frame, resulting in error accumulation frame by frame. If the tracking of a certain frame fails, the camera tracking cannot be continued, and the camera must be repositioned. Some methods propose to use random classifiers or Harrwaveltes to solve the camera relocation problem. Chekhlov et al. propose to use SIFT feature points to track feature points, but this leads to increased computational complexity. In order to reduce the computational complexity, some methods simplify th...

AI Technical Summary

Problems solved by technology

[0003] The difficulty of camera tracking technology lies in: 1) Most methods track feature points frame by frame, resulting in error accumulation frame by frame

If the tracking of a certain frame fails, then the camera will not be able to continue tracking, and the camera must be repositioned

Some methods propose to use random classifiers or Harrwaveltes to solve the problem of camera relocation. Chekhlov et al. proposed to use SIFT feature points to track feature points, but this leads to increased computational complexity.

In order to reduce the computational complexity, some methods simplify the SIFT detection algorithm, but then lose some invariant features of SIFT feature points; 2) most of the existing camera tracking algorithms can only deal with static scenes, If the scene changes dynamically, it will cause the tracking to fail

However, real natural scenes are often changing, and the scenes that need to be processed in many applications are also dynamically changing, so the existing camera tracking methods are difficult to deal with these scenes

3) During the tracking process, some existing methods do not update and optimize the 3D scene structure in time, resulting in the reconstructed 3D model becoming invalid soon, which in turn leads to the failure of camera tracking

The methods proposed by Blaser and Shimamura can both guarantee the compactness and accuracy of the scene, but if there is a large occlusion, the standard RANSAC method will soon be unable to accurately estimate the camera parameters. In addition, vSLAM does not delete invalid keyframes, so it has been The 3D points of the new object blocked by the existing invalid key frame can no longer be added to the scene

Images