A key problem in object recognition is finding a suitable object representation. For historical and computational reasons, vector descriptions that encode particular statistical properties of the data have been broadly applied. However, employing tensor representation can describe the interactions of multiple factors inherent to image formation. One of the most convenient uses for tensors is to represent complex objects in order to build a discriminative description. Thus thesis has several main contributions, focusing on visual data detection (e.g. of heads or pedestrians) and classification (e.g. of head or human body orientation) in still images and on machine learning techniques to analyse tensor data. These applications are among the most studied in computer vision and are typically formulated as binary or multi-class classification problems. The applicative context of this thesis is the video surveillance, where classification and detection tasks can be very hard, due to the scarce resolution and the noise characterising sensor data. Therefore, the main goal in that context is to design algorithms that can characterise different objects of interest, especially when immersed in a cluttered background and captured at low resolution. In the different amount of machine learning approaches, the ensemble-of-classifiers demonstrated to reach excellent classification accuracy, good generalisation ability, and robustness of noisy data. For these reasons, some approaches in that class have been adopted as basic machine classification frameworks to build robust classifiers and detectors. Moreover, also kernel machines has been exploited for classification purposes, since they represent a natural learning framework for tensors.
|Titolo:||Tensor Representations for Object Classification and Detection|
|Data di pubblicazione:||2012|
|Appare nelle tipologie:||07.13 Doctoral Thesis|