Present position: IPSC Grantholder at the European Joint Research Center (JRC)
|Thesis title:||Structure Deformations and Projection Deformations in 3D Reconstruction|
|Research area:||Computer Vision|
The linear projection model used to describe an image sensor has been widely adopted in order to perform reconstruction of 3D scenes. It finds application in Mobile Robotics, as it can improve navigational capabilities, in Computer Graphics, as it automatically provides a 3D model of the environment, and in Video Surveillance, as it can enhance the automatic detection of targets.
Nevertheless this model requires a rigid environment and a standard perspective camera. The first assumption may not hold, in particular for applications dealing with video sequences of deforming, or independently moving, objects. In this case, infact, although linear relations exist for each single frame, a non-linear or linearized model describing the deformations must be adopted in order to perform 3D reconstruction. The second assumption may
not hold either, in particular in Mobile Robotics application where usually robots are equipped with special cameras such as omnidirectional system, in order to increase their field of view. For these special cameras the image formation model depends on their intrinsic properties, and thus non linearities are introduced directly by the projection sensor.
This thesis deals with both sources of deformations, i.e. deformations induced either by the projective system or by temporal changes in non-rigid objects.
These deformations introduce non linear relations among imaged pixels in each frame and 3D points in the image formation model. We show which information about the environment and the imaging sensor could be extracted from the perceived images sequence. Notice that the effects perceived by the sensor, and equivalently by the human eye, depend on both type of deformations.
We present some original contributions that deal with specific type of deformations.
We show how to exploit a rigid environment, and in particular the presence of straight lines, perceived by particular non linear projection systems to recover the camera calibration parameters. A second contribution deals with isometric deformations of paper-like surfaces seen under a standard perspective projection and shows how to perform 3D reconstruction by describing the underlining physical model of the deformation. Finally the last contribution uses techniques usually exploited for structure deformation in order to recover information about a planar motion performed by a general non-linear projective model.