Curve and surface reconstruction from sparse data has been recognized as an important problem in computer graphics. Non structured data points (i.e., a set of points with no knowledge of connectivity and proximity) together with the existence of noise make this problem quite difficult. In order to solve it, several techniques have been proposed, such as, some of them are based on Delaunay triangulation, other are based on implicit surface reconstruction or on the advancing front techniques. Our algorithm consists basically in four steps. In the first step, a clustering procedure is performed in order to group the sample points according to their spatial location. This procedure obtains an spatial structure for the points by subdividing uniformly the plane in rectangular cells, and classifying them into two categories: empty (when the cell contains no point inside) or not empty (otherwise). At this stage, a data structure is built in such way that it is possible to query the set of sample points that belong to a given rectangular cell. The second step processes the point through the Moving Least Squares method. Its objective is not only to reduce the noise on the data, but also to adapt the number of point to the desired level, by adding or removing points from the initial set. The third step builds the skeleton of the set of cells that have sample point on its interior. Such skeleton is in fact a digital approximation for the curve that will be reconstructed. It is obtained by the use of a topological thinning algorithm, and its implementation is done in such a way that the number of points in each cell is considered, for example, the cells with less number of points are not considered for the thinning. In the last step, the curve is finally reconstructed To do so, the skeleton obtained in the third step is used to construct a piecewise-linear approximation for the curve, where each vertex is obtained from the MLS projection on the middle point of the skeleton rectangular cell.