A Stochastic Approach to Anomaly Detection and Classification in Multidimensional Time Series

Central in this works is to perform anomaly detection on video signals from CCTV cameras in search of dangers to the public safety, which then become the anomaly in this context. A framework is therefore devised, based on an interplay between Convolutional Neural Networks and Hidden Markov Models. The system obtained is tested both against the theoretical results found and on the task of identifying car crashes on real videos.