光流特征聚类的车载全景序列影像匹配方法

An Image Match Method Based on Optical Flow Feature Clustering for Vehicleborne Panoramic Image Sequence

提出一种光流特征聚类的车载全景序列影像匹配方法。采用非参数化的均值漂移特征聚类思想,以SIFT多尺度特征匹配点的位置量和光流矢量,构建了影像特征空间的空域和值域;利用特征空间中对应的显著图像光流特征为聚类条件,实现了全景序列影像的匹配;最后以全景极线几何约束为条件进行粗差的剔除。通过相同、不同内点率以及不同数据的试验对比分析,本文方法在匹配正确点数和正确率方面要优于经典的Ransac法和金字塔Lucas-Kanade光流法,尤其在场景复杂造成的低内点率情况下,算法表现较为稳定,并可较好地剔除由重复纹理、运动物体、尺度变化等产生的匹配点粗差。

An image match method based on optical flow feature clustering is presented for vehicle-borne panoramic image sequence.The spatial domain and range domain of image feature space are built by the coordinates of SIFT multi-scale feature matching point and optical flow vector,then the panoramic image match is finished by Mean Shift attached to the optical flow clustering constraint condition in image feature space.Finally,panoramic geometric constraint of Ransac method is used for gross error detection.Several panoramic images are selected and used for experiment.The experiments of analysis and comparison were carried out in the conditions of the same inlier ratio,different inlier ratio and different data.The results show that the proposed method in the number and accuracy of correct matching points are superior to classic Ransac method and Pyramid Lucas-Kanade method,especially in the complex scene in low inlier ratio cases,the algorithm performance is relatively stable,and have better constraint effect for the gross error usually caused by repeat texture,moving objects and scale change.