Local feature extraction plays an important role in related point cloud applications. Therefore, an EPFH (extended point feature histograms) descriptor for the local feature representation of 3D point cloud was proposed. Each point pair was represented by several invariant pairwise point attributes. Then, a local reference frame was defined for a keypoint and the neighboring points of the keypoint were transformed into the local reference frame. These pairwise points attributing between the neighboring points and the keypoint were accumulated into several subfeatures in a set of subspaces. These subfeatures were finally concatenated and compressed into an overall feature descriptor. The EPFH descriptor was tested by a popular publicly available Bologna dataset and was compared with several existing methods. Experimental results show that the proposed EPFH method outperforms several existing methods under different levels of noise and point cloud resolutions.