Abstract: Normal vector estimation in three-dimensional space is of great significance in the field of research in computer vision and surface reconstruction, the local surface fitting method is a classical estimation method of point cloud data. In order to improve the veracity of the normal vector which computed by the way of local surface fitting, a method based on optimal neighborhood size for normals estimation was described and analyzed in this paper. The distribution of the neighbor of every point was formulated on the basis of the projection of gradient. Then, the adaptive size was chosen based on the distribution of the neighbor, the normal vector was fitted by the adaptive size. Experimental results show that presented algorithm could avoid the radius of neighbor estimated too large or too small, improve the veracity of the normal vector which computed by the way of local surface fitting effectively.