Abstract :
[en] Recent advances in commodity depth or 3D sensing technologies have enabled us to move
closer to the goal of accurately sensing and modeling the 3D representations of complex
dynamic scenes. Indeed, in domains such as virtual reality, security, surveillance and
e-health, there is now a greater demand for aff ordable and flexible vision systems which
are capable of acquiring high quality 3D reconstructions. Available commodity RGB-D
cameras, though easily accessible, have limited fi eld-of-view, and acquire noisy and low-resolution measurements which restricts their direct usage in building such vision systems.
This thesis targets these limitations and builds approaches around commodity 3D
sensing technologies to acquire noise-free and feature preserving full 3D reconstructions
of dynamic scenes containing, static or moving, rigid or non-rigid objects. A mono-view
system based on a single RGB-D camera is incapable of acquiring full 360 degrees 3D reconstruction of a dynamic scene instantaneously. For this purpose, a multi-view system
composed of several RGB-D cameras covering the whole scene is used. In the first part of
this thesis, the domain of correctly aligning the information acquired from RGB-D cameras
in a multi-view system to provide full and textured 3D reconstructions of dynamic
scenes, instantaneously, is explored. This is achieved by solving the extrinsic calibration
problem. This thesis proposes an extrinsic calibration framework which uses the 2D
photometric and 3D geometric information, acquired with RGB-D cameras, according
to their relative (in)accuracies, a ffected by the presence of noise, in a single weighted
bi-objective optimization. An iterative scheme is also proposed, which estimates the parameters
of noise model aff ecting both 2D and 3D measurements, and solves the extrinsic
calibration problem simultaneously. Results show improvement in calibration accuracy
as compared to state-of-art methods. In the second part of this thesis, the domain
of enhancement of noisy and low-resolution 3D data acquired with commodity RGB-D
cameras in both mono-view and multi-view systems is explored. This thesis extends
the state-of-art in mono-view template-free recursive 3D data enhancement which targets
dynamic scenes containing rigid-objects, and thus requires tracking only the global
motions of those objects for view-dependent surface representation and fi ltering. This
thesis proposes to target dynamic scenes containing non-rigid objects which introduces
the complex requirements of tracking relatively large local motions and maintaining data
organization for view-dependent surface representation. The proposed method is shown
to be e ffective in handling non-rigid objects of changing topologies. Building upon the
previous work, this thesis overcomes the requirement of data organization by proposing
an approach based on view-independent surface representation. View-independence
decreases the complexity of the proposed algorithm and allows it the flexibility to process
and enhance noisy data, acquired with multiple cameras in a multi-view system,
simultaneously. Moreover, qualitative and quantitative experimental analysis shows this
method to be more accurate in removing noise to produce enhanced 3D reconstructions
of non-rigid objects. Although, extending this method to a multi-view system would
allow for obtaining instantaneous enhanced full 360 degrees 3D reconstructions of non-rigid
objects, it still lacks the ability to explicitly handle low-resolution data. Therefore, this
thesis proposes a novel recursive dynamic multi-frame 3D super-resolution algorithm
together with a novel 3D bilateral total variation regularization to filter out the noise,
recover details and enhance the resolution of data acquired from commodity cameras in
a multi-view system. Results show that this method is able to build accurate, smooth
and feature preserving full 360 degrees 3D reconstructions of the dynamic scenes containing
non-rigid objects.
