Due to the searching tasks of the small and/or color objects on the seabed, the multi-AUV systems with the optical sensors have become a research hotspot. To build a given pattern formation(a planar pyramid pattern) and put all the small homogeneous AUV units together, based on the relative positions got from the optical sensors and the absolute orientation evaluated by the compass, a local position-based control method is proposed. This method includes two parts: 1) a neighbor-check mechanism is given to distinguish the AUVs' IDs; 2) a collision avoidance strategy with complexity $\mathcalO$(nlogn) is proposed to optimize the position and gesture of a planar pyramid pattern, and the non-intercrossing linear trajectories for each AUV is planed. The proposed control method is tested in a realistic obstacle-free deep-sea simulation environment established in Blender. 4sim7 AUVs(CISCREA) are used to build the planar pyramid pattern repeatedly. The results verify the feasibility and stability of the proposed control method.