Triangular surface reconstruction out-of-point clouds suffer from noisy, non-uniform distributed data, and complicated topology structure. Thus, an improved growing neural gas approach is proposed. A point cloud projection on local grid is employed to direct node insertion; therefore, to adaptively control neuron growing rate, the geometric and topologic transforms are sychronized. Redundant links are removed through non-manifold edge detection, that guarantees a topologically validate mesh. The network keeps updating triangular grid and then fills holes in a post phase by the extended neighborhood connection mechanism. After all those steps come to a convergent end, there is a gap free and an Euler characteristic correct mesh was obtained. Case studies invalidate the noise robustness and complex topology adaptability. The algorithm cand further adjust mesh size to point cloud distribution. Plus is that reconstructed mesh approximates the surface in high accuracy, and it characterizes uniform equilateral edge share.