Building extraction plays an important role in building reconstruction and urban management. In this study, a normalized difference vegetation index (NDVI) constrained watershed segmentation algorithm is utilized to segment airborne LiDAR data, and certain criteria are used to discriminate building regions as follows. First, grid data is attained by the interpolation of LiDAR point clouds. Then, the NDVI constrained watershed segmentation algorithm is applied to segmenting the digital surface model data, which is generated from LiDAR. Further, NDVI is introduced in the flooding process of the watershed algorithm to separate the vegetation from the buildings. Finally, the building regions are identified through some of the criteria (elevation difference, size, and NDVI) according to the adjacency relationship of each region. The benchmark data of the International Society for Photogrammetry and Remote Sensing for Vaihingen are used to evaluate the building detection results. The average completeness, correctness, and quality are respectively 89.2%, 94.3%, and 84.7% at the pixel level and 81.8%, 93.1%, and 76.9% respectively at the object level. Moreover, for an object with area larger than 50 m2, the average completeness, correctness, and quality are 99.1%, 100%, and 99.1%, respectively.