Master Assignment
RPi-based passive 5-camera system for 3d face acquisition
Type: Master M-EE
Location: University of Twente
Duration: Nov, 2018 - Jul, 2019
Student: Valk, D.J. van der (Diederik, Student M-EE)
Date Final project: July 15, 2019
Supervisors:
Abstract:
A Raspberry Pi (RPi-) based 5-camera system was analysed for its potential to create high resolution - up to 0.2mm depth steps - 3D reconstructions of faces. A theoretical framework to determine the limits of the system was proposed, including the input depth resolution, a precision qualifier and maximum time synchronisation offsets between the cameras. The theoretical limits however could not (yet) be verified in practice. The found limits for this specific implementation were a max distance of 14cm from the cameras for a 0.2mm input depth resolution, and a theoretical maximum time offset of 0.016ms. A practical methodology was set up to generate a 3D face reconstruction starting from lens choices and adjustments, via calibration to reconstruction and comparison with other references. This resulted in a reconstruction at 0.4m where > 90% of the 3D reconstructed points were max 3mm off compared to a reference, set by a medical 3D face reconstruction device - the 3dMD. Various prospects for further improvement were found.
Introduction:
Biometrics is about recognizing persons based on their physical properties, behavior or traces they leave. Examples of biometric modalities are face, fingerprint, iris, voice etc. An interesting modality for face recognition is the 3D shape of the face. In the DMB group, we have developed very good algorithms for 3D face recognition that are among the best in the world. However, there is a lack of cheap, accurate 3D cameras. In addition, most 3D cameras are active cameras that project a pattern on the objects (faces) and cannot be used well in bright day light.
Figure 1: From left to right: 5 camera setup front and back, 3d face surface and range image
Assignment:
In order to investigate the feasibility of a passive 3D camera at DMB we built a setup with 5 cameras driven by 5 Rasberry Pies (RPI). In the past, we already developed basic methods for reconstruction of 3D objects from the images of 5 cameras. This assignment involves porting the software to the 5 camera RPI platform, calibrate the camera setup and optionally, further optimise the software and hardware. Furthermore, 3D facial images are to be recorded and it should be investigated how well they can be used together with our 3D face recognition software.