基于特征点跟踪的视频序列平面提取算法

Multiple Plane Extraction Based on Feature Point Tracking over Video Sequences

平面结构是一种最为常见和重要的几何特征,平面提取在很多领域有着重要的应用。提出了一种毋须恢复摄像机三维信息、直接从视频序列中提取出平面的算法。先通过在视频相邻帧之间建立极线约束下的射影变换模型,来计算出由平面诱导的单应,而每一个单应都对应着场景中的一个平面;然后根据计算出的单应对视频的每一帧进行分割,进一步提取出视频中的平面结构。一系列真实视频序列的实验验证了本算法的有效性和稳定性。

Planar structures are abundant in both man-made and natural environments which enables the use of planes in various vision tasks. We introduced a new approach for the robust detection of multiple planes over a video sequence without camera calibration or 3D reconstruction. Given a video sequence, we first built a projective geometry model between adjacent frames based on epipolar constraint. Then a homography induced by a plane was computed, with which we could get the planar structure by doing video segmentation. Experimental results on a variety of real video sequences have verified the effectiveness and efficiency of our method.