As the three-dimensional scanning technology has advanced rapidly in the recent several decades, it gets to draw growing interest in shape modeling from scanned pointcloud data to satisfy the increasing need for reverse engineering of physical artifacts in aerospace, automotive industry, customization-oriented product design and manufacturing. For innovative product design, the three-dimensional (3D) scanning is a convincing technology to convert a physical model into point cloud digital one, which allows quickly reuse of those interesting shape features from virtually all existing objects. As one typical output format, three-dimensional point cloud data (PCD) has been applied more and more frequently in product design procedures. However, the PCD without pre-process cannot be used directly for three reasons: 1. The number of points is so large that the post-process time proportionally increases; 2. The initial PCD contains noisy points and incomplete information; 3. The points are scattered and irregular, increasing the error rate of post-analysis process. In order to solve such vi problems existing in the raw point cloud data set obtained by scanning, we propose a framework on self-adaptive k-means clustering algorithm and data reduction method for 3D point cloud data processing and design reuse to generate a satisfactory 3D model for product design and manufacturing. With the development of 3D scanning technology and devices, the PCD has become more accurate while with more points. When handling the massive data which contains huge amount of points and is embedded with inevitable noise, it makes the data processing, mainly by means of triangulated irregular networks through automated procedures, more sophisticated. So the capability to cluster the point cloud data into a set of clusters is critical for the downstream procedures including model feature extraction and innovative design. The conventional k-means clustering algorithm is well adapted to this problem. However, as it is sensitive to the empirically specified initial cluster centroids and the value of clusters, k-means method could only be used to manipulate the two-dimensional small data. It tends to be unaffordable in terms of computational complexity when dealing with three-dimensional point cloud big data problem. In this paper, a self-adaptive cluster selection mechanism is proposed to improve the k-means algorithm. The number of clusters is estimated based on the distribution of local features, and the optimal initial cluster centroids is computed based on randomized search and similarity measure of area density. After clustering the PCD into a certain number of clusters, de-noising is operated to filter all the outliers to make the data reduction more accurate. In this paper, the outliers are classified into three categories: non-isolated outliers, isolated outliers and scattered noisy points and three different methods are proposed to filter these outliers respectively. First a majority voting algorithm is adopted to detect the non-isolated outliers and cut the connection between them and the needed surface point set, changing the non-isolated outliers to isolated ones. Then the isolated outliers are removed based on an expandable boundary criterion. At last the scattered noisy points are removed based on a novel filter based on the integration of k-means clustering algorithm and vii moving least squares (MLS) method. Then a novel data reduction and regularity framework for scattered PCD is proposed in this paper. This approach adopted an feature-detective filter to remove the sparse noisy points and a local feature approximation based cave-filling method to fill the incomplete information, then proposed an innovative planar sections based algorithm to transfer the threedimensional point cloud data reduction and regularity problem into a two-dimensional problem and obtain a set of suitable parameters through a method based on particle swarm optimization (PSO) algorithm. Whilst a similarity assessment method based on Distance Shape Histogram (DSH) is adopted to verify the efficiency of the proposed method and its ability to keep high similarity with the original data set. The method has proved its effectiveness with plenty of examples. Experiments are performed to evaluate the effectiveness and efficiency of the proposed method using synthetic data and real data and the results demonstrate that the proposed method can generate an appropriate 3D models for product design and reuse. Keywords— point cloud data, design reuse, k-means, clustering, reduction and regularity, moving least squares(MLS), principal component analysis(PCA)