车载组合全景相机单片影像匹配算法

摘要

车载组合全景影像存在严重的射影变形,纹理单一重复,造成全景单片影像匹配同名点的正确率低.由于纹理重复方向、运动物体的移动方向与摄影基线方向一致,用随机抽样一致性算法(RANSAC)结合极线几何不能有效地剔除此类误匹配,该文建议一种适宜车载组合全景单片影像的匹配流程.首先提取Harris角点特征,SIFT描述符描述特征和距离比阈值约束匹配初始同名点;然后,RANSAC结合极线几何初步剔除误匹配,特征点追踪获取同名点的多视对应,利用GPS和IMU提供的影像外参数多片前交剔除误匹配;最终通过地面点的反投影像点模板匹配,进一步增加同名像点的重叠范围.据实验分析,本流程匹配到的同名点不含有粗差,且数量充足,可满足后续空中三角测量的需要.%Vehicle-mounted composite panoramic images have the problems of serious projective deformation,simple and repeat texture,which cause low matching accuracy among images.Because the texture repeat direction and moving direction of the moving object have the same direction with photography baseline,RANSAC combined with the polar geometry cannot effectively eliminate such mismatches.This paper suggests a suitable matching process for vehicle-mounted composite panoramic images.Firstly,extract Harris corner features,and describe features with SIFT descriptors.Then,match features with threshold of distance ratio,delete primary mismatches using RANSAC combined with the polar geometry,track matches to get all image points corresponding to the same ground point,and delete final mismatches using multi-image fore-intersection with extrinsic parameters provided by GPS and IMU.Finally,further increase the overlap of homologue points using template matching with the back projection image points of the ground point.According to the experimental analysis,this process gets homologue points without any gross error,and these points are abundant and can meet the needs of the subsequent aerial triangulation.