Abstract：This paper proposes an improved patch-based multi-view stereo(PMVS) algorithm,aiming to overcome its limitations in the dense matching of multi-view oblique images.The new algorithm combines with the objects features in a three-dimensional modeling method for cities,makes use of elevation constrains,cluster analysis and candidate image sorting strategies,and it also uses a grid diffusion to substitute the original hexagonal-direction diffusion,and suits to large oblique images.An experiment shows that this improved PMVS can effectively restrict the seed number in initial matching,and greatly improve the accuracy and quality of seed points.The subsequent diffusion and filtering uncertainty are reduced.With this new algorithm,the final number of seed points increases 78%,the point cloud vacuums decrease significantly,even totally vanish in some cases.It provides a new technology for DEM production and a three-dimensional modeling method for cities.

Keyword：multi-view image;dense matching;patch-based multi-view stereo algorithm;points cloud;