This paper proposes a method of digital surface model (DSM) generation based on the multiple-aerial-image matching, which can comprehensively make use of the information from the image- and object- space. Firstly, through matching between the reference image and each slave image, any possible candidate in each stereo pair can be considered, and this guarantees the full use of the image photometrical information; then the matching results from all stereo pairs are merged in the object space, which can partially detect and eliminate mismatching points in each stereo pair, to provide more accurate spatial information to restrict and guide the matching in the next pyramid level. In the process of matching, the geometrically constrained normalized cross-correlation matching and global relaxation optimization strategy are incorporated. Some experiments were made on a set of large overlap aerial digital frame imagery of some area to validate the method. The results have shown that comparing to the accuracy of DSM generated by two images, the accuracy by three images has been improved evidently, but along with the number of the slave image increases, the matching success ratio as well as the accuracy of the DSM will not again have the remarkable improvement.