Online control of active camera networks for computer vision tasks
ACM Transactions on Sensor Networks (TOSN)
| Hi-index | 0.00 |
Modeling dynamic scenes/events from multiple fixed-location vision sensors, such as video camcorders, infrared cameras, Time-of-Flight sensors etc, is of broad interest in computer vision, with many applications including 3D TV, virtual reality, medical surgery, markerless motion capture, video games, and security surveillance. However, most of the existing multi-view systems are set up in strictly controlled indoor environments, with fixed lighting conditions and simple background views. Many complications limit the technology in the natural outdoor environments. These include varying sunlight, shadows, reflections, background motion and visual occlusion etc. In this thesis, I address different approaches overcoming all of the aforementioned difficulties, so as to reduce human preparation and manipulation, and to make a robust outdoor system as automatic as possible.The main novel technical contributions of this thesis are as follows: a generic heterogeneous sensor fusion framework for robust 3D shape estimation; a way to automatically recover 3D shapes of static occluder from dynamic object silhouette cues, which explicitly models the static "visual occluding event" along the viewing rays; a system to model multiple dynamic objects shapes and track their identities simultaneously, which explicitly models the "inter-occluding event" between dynamic objects; and a scheme to recover an object's dense 3D motion flow over time, without assuming any prior knowledge of the underlying structure of the dynamic object being modeled, which helps to enforce temporal consistency of natural motions and initializes more advanced shape learning and motion analysis. A unified automatic calibration algorithm for the heterogeneous network of conventional cameras/camcorders and new Time-of-Flight sensors is also proposed.