Identificazione di anomalie nell’attenzione del guidatore e nel comportamento delle persone.

Abstract: As the world matures increasingly connected and digitized by the day, with sensors and computing devices becoming more and more pervasive, new opportunities appear for artificial intelligence. In particular, public monitoring steps forward as a critical theme, and computer vision can forcefully prevail as the lead technology to help build a safer world. In this thesis, we present solutions to tackle public safeguard in two different areas of operation. First, we begin with vehicle-based safety by developing a system capable of predicting where a person is likely to focus her attention on while driving. Such activity has a vast potential to improve driving safety. Nevertheless, it appears utterly complex since driving a car is a complicated task, and it is highly subjective from an attentive perspective. To handle attention prediction, we collect and release DR(eye)VE, a dataset consisting of driver-centric and car-centric clips, along with driver's fixation points on the outer urban scene. Next, we deeply inspect such data in order to establish which factors most influence a driver's gaze, both in terms of motion and semantics. Guided by such evidence, we finally develop a deep neural network that, given a car-centric urban scene, identifies which regions are likely to capture the driver's attention. Secondly, we address surveillance-based safety by introducing an anomaly detection model capable of learning the traits that characterize healthy (safe) situations and, therefore, alert when unexpected events appear. Learning such models without utilizing examples of abnormal conditions is the aim of anomaly detection (a.k.a. novelty detection) research. Despite its importance and a plethora of prior work, the unpredictable nature of novel events and their inaccessibility during the training procedure severely degrades the effectiveness of state-of-the-art systems. In this framework, we propose a general model consisting of a deep autoencoder equipped with a parametric density estimator, fitting its latent representations through an autoregressive procedure. We show that a maximum likelihood objective in latent space effectively regularizes the optimization of the autoencoder's reconstruction error, and minimizes the differential entropy of the distribution spanned by latent vectors. Intuitively, such a joint optimization forces the model to describe (and reconstruct) each example in terms of features that frequently appear in the training set. Extensive experimental inquiries and comparisons with prior art show the effectiveness of both our proposals.

 not available