Using unmanned aerial vehicle (UAV) images to inventory forest resource is a quick solution to collect high resolution data with rich imagery details. It is capable to recognize individual trees with locations and crown sizes. An intrinsic problem of high spatial resolution UAV images at centimeter levels is that the images are tended to oversegmented. In addition, UAV images captured in plant growing season can hardly observe the ground and objects beneath the canopy top, leading to infeasibility of height normalized canopy height model (CHM) based crown segmentation algorithms in forested areas with large terrain variations. To tackle these problems, a novel UAV image crown extraction approach was proposed, which was free of height normalization. Firstly, a 3D surface model was built from dense images by structure from motion technology. Initial tree locations were identified by combining height information and image contexts. An adaptive kNN neighborhood watershed algorithm was implemented to derive crown coverage of each initial tree locations. UAV images of Larch forests in Baihuashan National Nature Reserve of Beijing were used to conduct the experiment, and it was validated by visual interpretation on orthophotos and compared with a couple of images or point cloud based automatic segmentation algorithms. The results showed that the overall detection rate of individual trees was over 91%. The crown size extraction accuracy was over 81%, which outperformed the original watershed and other crown segmentation methods. It was demonstrated that the proposed method can serve to extract high accuracy tree parameters rapidly at large scales in complex terrain environment.