Titolo della tesi: Exploiting local features in modern Computer Vision
Computer Vision is an interdisciplinary field which refers to replicating the human
vision system and enabling machines to gain high-level understanding from digital
images or videos. It plays an important role in many applications, which have
become commonplace today such as face recognition, anomaly detection, behaviour
recognition, re-identification, image segmentation, and so on.
During the last years, the widespread of new visual sensors (i.e., Pan Tilt Zoom
camera, depth cameras, acoustic sensors, etc.) have enlarged the number of possible
computer vision applications. Particular attention has been focused on those applications
in which unconstrained still images and videos have to be processed. This
has raised at the same time new research challenges. Among the most interesting
topics one which is receiving a growing interest is the identification and extraction
of distinct features.
The local features are one of the most used features in the current literature. They
allow to detect and extract features, which are robust to several types of transformation
or distortion (i.e., occlusion, rotation, scale). However, this type of features
has been mainly used in RGB camera-based applications.
The goal of this thesis is presenting and demonstrating the efficacy of novel computer
vision research contributions to different application scenarios as background modelling,
RGB-D data fusion for event detection and underwater image compression
and transmission. In these applications, we will show how local features can be used
on unconstrained environment and with image and videos acquired from different
visual sensors.