Human detection is a basic functionality for home service robots. For complex family environments where lying person is partially occluded or in cluster, this paper proposes a lying person detection approach integrates 3D point cloud segmentation and local feature matching. Our approach segments the point cloud of each object into several pieces, matches local features of each object piece, and classifies them to detect lying person. Experiments show that our approach can achieve high detection accuracy with average recognition time less than 0.3s. Our approach meets the human detection requirements for service robots and is demonstrated to be practical and reliable, even when parts of human body is occluded.