Depth sensors of the type RGB-D are interesting alternatives to do a 3D reconstruction of the environment with low cost. In this work, we evaluate a reconstruction pipe-line implemented on GPU that merges a point cloud alignment tracking system to estimate the camera position and a volumetric reconstruction/visualization to smooth the naturally noise measures from the sensor, using the Kinect as RGB-D input. A technical analysis of the algorithm has been made, showing the impact of parameter modifications of the system and a comparative of time and precision between the presented implementation and a public version available by the Point Cloud Library (PCL) during the development of this work. Some modifications to the original work have been made and the performance tests demonstrate that our implementation is faster than the PCL version without compromising the reconstruction precision.