DBSCAN modificado con Octrees para agrupar nubes de puntos en tiempo real

With the advent of commercial depth sensors, such as the Kinect, new opportunities to develop applications and interactive systems that use the human body are created. However, those sensors capture a large amount of 3D data, known as point clouds, that have to be processed, trying to obtain as much relevant information as possible. Clustering algorithms, normally used in data mining, are useful to 173 Research in Computing Science 114 (2016) pp. 173–186; rec. 2016-03-09; acc. 2016-05-08 discover that information. One of the most known is DBSCAN, which allows the creation of an unknown number of clusters within a data set, while filtering out noise. Unfortunately, the algorithm can also be slow due to the data type of the data set, as well as because of the search of data with similar characteristics. In this work, the use of the DBSCAN algorithm to process point clouds is proposed. Additionally, a modification of the algorithm is presented: using octrees to decrease the neighbor searching times, and using a data partitioning scheme in order to reduce the number of elements of each search. With the proposed method, the point cloud processing time was considerably accelerated when compared to the original algorithm, achieving real time processing, while obtaining the same results.