A Surfels 3D reconstruction method based on improved KD-Tree is put forward, firstly collecting the discrete point cloud data through RGB-D camera, replacing the circular or oval surfel model with hexagonal model for modeling and determining the surfel radius in light of neighborhood distribution of sample points; Moreover, doing inside and outside relations test between one point model and another discrete point model, building KD-Tree for each model, setting the axis with the longest projection length as the separating axis, improving segmentation rules, accelerating the detection of inside and outside and intersecting relations. Experiments show this algorithm has great reconstruction effects on the 3D reconstruction both of heterogeneous sample points and discrete point cloud with different resolution with steady and efficient calculation.