3D modeling by low-cost range cameras: methods and potentialities

Nowadays the demand of 3D models for the documentation and visualization of objects and environments is continually increasing. However, the traditional 3D modeling techniques and systems (i.e. photogrammetry and laser scanners) can be very expensive and/or onerous, as they often need qualified technicians and specific post-processing phases. Thus, it is important to find new instruments, able to provide low-cost 3D data in real time and in a user-friendly way. Range cameras seem one of the most promising tools to achieve this goal: they are low-cost 3D scanners, able to easily collect dense point clouds at high frame rate, in a short range (few meters) from the imaged objects. Such sensors, though, still remain a relatively new 3D measurement technology, not yet exhaustively studied. Thus, it is essential to assess the metric quality of the depth data retrieved by these devices. This thesis is precisely included in this background: the aim is to evaluate the potentialities of range cameras for geomatic applications and to provide useful indications for their practical use. Therefore the three most popular and/or promising low-cost range cameras, namely the Microsoft Kinect v1, the Micorsoft Kinect v2 and the Occipital Structure Sensor, were firstly characterized from a geomatic point of view in order to assess the metric quality of the depth data retrieved by them. These investigations showed that such sensors present a depth precision and a depth accuracy in the range of some millimeters to few centimeters, depending both on the operational principle adopted by the single device (Structured Light or Time of Flight) and on the depth itself. On this basis, two different models were identified for precision and accuracy vs. depth: parabolic for the Structured Light (the Kinect v1 and the Structure Sensor) and linear for Time of Flight (the Kinect v2) sensors, respectively. Then the effectiveness of such accuracy models was demonstrated to be globally compliant with the found precision models for all of the three sensors. Furthermore, the proposed calibration model was validated for the Structure Sensor: with calibration, the overall RMSE, decreased from 27 to 16 mm. Finally four case studies were carried out in order to evaluate: • the performances of the Kinect v2 sensor for monitoring oscillatory motions (relevant for structural and/or industrial monitoring), demonstrating a good ability of the system to detect movements and displacements; • the integration feasibility of Kinect v2 with a classical stereo system, highlighting the need of an integration of range cameras into 3D classical photogrammetric systems especially to overpass limitations due to acquisition completeness; • the potentialities of the Structure Sensor for the 3D surveying of indoor environments, showing a more than sufficient accuracy for most applications; • the potentialities of the Structure Sensor to document archaeological small finds, where metric accuracy seems to be rather good while textured models shows some misalignments. In conclusion, although the experimental results demonstrated that range cameras have the capability to give good and encouraging results, the performances of traditional 3D modeling techniques in terms of accuracy and precision are still superior and must be preferred when the accuracy requirements are restrictive. But for a very wide and continuously increasing range of applications, when the required accuracy can be at the level from few millimeters (very close-range) to few centimeters, then range cameras can be a valuable alternative, especially when non expert users are involved. Furthermore, the technology on which these sensors are based is continually evolving, driven also by the new generation of AR/VR reality kits, and certainly also their geometric performances will soon improve.