With the aim to reconstruct three dimensional virtual scenes, a new algorithm for automatically generating three dimensional colored point clouds from aerial images is proposed in this paper. The algorithm adopts a new multi-view image probability relaxation global matching strategy based on fusion of object space and image space information to synthetically utilize multi-image information to improve matching reliability. Then, according to multi-view image matching results, multi-image bundle adjustment is used to compute the three dimensional coordinates of matching points. Finally, a statistical test on correctness of adjustment model is adopted to quantitatively measure the accuracy of computed three dimensional coordinates. The proposed algorithm is applied to actual aerial images. The experimental results on a statistical test indicate that matching reliability is above 90%, which means that the proposed algorithm has higher robustness and precision. Through proposed method, all planar pixels in the overlapping region of multi-view image are transformed to three dimensional points, and three dimensional colored point clouds with full digital format is automatically generated to represent geographic scenes. It can well satisfy the requirements of quick reconstruction of three dimensional geographic scenes with large scopes.