基于长序列图像的三维重建技术研究*

提出一种基于长序列未定标图像的三维重建方法,并将其成功地应用于增强现实实例中.首先,基于传统KLT跟踪算法提出了一种针对序列图像的改进特征点匹配策略,通过特征点运动向量的预测减小了相应特征点的搜索范围,进一步根据相近特征点邻域窗口在透视畸变上的相似性大大提高了匹配算法的效率;在得到序列图像的匹配结果后,将传统基于仿射成像模型的测量矩阵(Measurement Matrix)保秩分解算法扩展到透视成像模型中,从而一次性得到整个场景的射影重建;进一步在摄像机自定标的基础上得到整个场景的三维欧氏模型和摄像机的成像矩阵.最后给出真实图像序列的三维重建实验结果,并成功地将其应用到增强现实实例中.

3D Reconstruction Based on Long Image Sequence

Author proposed a 3D reconstruction method based on long uncalibrated image sequence,and the results were used in augmented reality applications successfully.First,an improved version of traditional KLT tracking method was proposed,based on the motion estimation of corresponding feature point and the similarity of geometric deformations of near feature window around feature point,which strongly improved the efficiency of matching algorithm.After getting the matching results across the whole sequence,an extension of rank fitting decomposition method for the measurement matrix used in affine model to perspective case was used to get the projective reconstruction.Then,a self-calibration was carried out for Euclidean reconstruction of the scene.Results of real image sequence and its applications for augmented reality demonstrate the feasibility of our method.