Research On The Method For3D Recognition For The Target Based On Image Sequence

3D reconstruction for the target based on monocular image sequence is one frontierand hotspot technologies of the computer vision. Compared to laser scanningtechnology based on active vision, which has simple equipment, flexible operation,low-cost, practical environment strong etc features. It has broad prospect ofapplication and can be used for target identification, weapons and navigation, aerialsurvey, war and natural disaster assessment, virtual reality and other fields. This paperfocuses attention on the research is how to obtain the object of image sequence andimage matching accuracy and efficiency. The main work is as follows:1. Introduced the theoretical basis of the three-dimensional reconstruction based onthe goals of image sequences, including the coordinates, camera model, epipolargeometry, fundamental matrix, dense disparity map and depth chart, etc.2. On the basis of the current several popular feature point extraction and matchingmethods. This paper presents an improved fast feature matching based on SURF: first,introduction to the NCC matching method for SURF, so that not only accelerate thespeed of the SURF matching, but also improve the accuracy of matching; then, themismatching points are excluded by RANSAC; Finally, using the least squares methodto fitting a straight line for the remainder of the match points, and obtained precision,robustness matching points, which provide a good foundation for follow-up accurateseed points of dense matching.3. In order to obtain dense3D point clouds of the target based on image sequence,it is necessary to ensure high accuracy of the target extraction and the disparity of thedense matching based on the target based on image sequence is denseness. To this end,this paper presents an improved target detection and dense matching method: usingmany interactive iteration segmentations method for the image object extraction tomake GrabCut algorithm segmentation obtained inaccurate area to be repaired; usingthe image texture features for the dense matching to divide image into the texture-richregion and sparse texture region. This paper takes different dense matching method tosolve the problem, which is the traditional dense matching based on region growing cannot be achieved quasi-dense matching in sparse texture region. The one hand, thismethod can solve the problem of inaccurate target extraction in image from complexscene; On the other hand, this method can solve the problem of the image sparse texture region can not be achieved quasi-dense matching and be resulted in the localizedreconstruction. The experimental results show that the results are better in the imageobject extraction and image dense matching, ultimately, this method improve theaccuracy of three-dimensional reconstruction.4. Design of3D reconstruction system based on the target of image sequence. Thispaper designs the whole system work flow and the various functional modules of thesystem, and discussed many problems, discussed camera motion parameters estimatedfrom the follow-up image matching,3D points coordinate calculation, triangulation andthe principle of texture mapping. Finally, developed a3D reconstruction software basedon the target of image sequences, besides, checked the feasibility and correctness of thismethod from the results of experiments.