|Thesis abstract: |
The Internet of Things is leading in the long term vision to a new paradigm in which low cost and energy aware networked sensing nodes are capable of carrying out complex visual analysis tasks.
The traditional paradigm for such tasks follows a "compress-then-analyze" model, as the sensing nodes are only responsible for the acquisition of the visual content from the environment and its efficient transmission to a central processing node. This methodology results in undue wastes of computational resources, since a part of the visual data might not be needed for the execution of the analysis tasks but is indeed transmitted to the central unit. The continuous technological evolution is leading to the production of more and more powerful yet energy-aware devices, allowing for the introduction of a new paradigm, namely "analyze-then-compress". According to such a model, the nodes are responsible not only for data acquisition and transmission but also for on-board analysis.
Visual Features represent a centerpiece for visual analysis, constituting a generic yet versatile tool for a large number of heterogeneous visual analysis tasks, such as object recognition, identification and tracking, pose estimation, face recognition, etc.
Resorting to Visual Features, some tools, methodologies and techniques are to be introduced to efficiently implement the "analyze-then-compress" paradigm on heterogeneous networks. Firstly, methods for Visual Feature extraction are to be tailored to the specific scenario of an energy aware network, then some data compression techniques are to be introduced to support an efficient transmission of the information over such a network. Cooperation between nodes will be taken into strong consideration to reduce the waste of resources, to enforce a fair division of the workload and to improve the performance of each task.