Learning to Classify Visual Dynamic Cues Learning to Classify Visual Dynamic Cues Title: Learning to Classify Visual Dynamic Cues

Classification based on dynamic information is a challenging research domain that finds application in a number of fields, including video-surveillance and video retrieval. Traditional approaches based on motion analysis address many interesting applications, such as access control, anomaly detection, congestion analysis and multi-camera event description: in all these cases it is common practice to devise a measurement phase that extracts low level information from videos. To this purpose a wide variety of methods have been presented in the computer vision literature, leading to solutions that effectively describe the video content in moderately difficult conditions. A well known limit of these methods is that while they provide effective tools to model the dynamics of a single video, they do not suffice when the problem of interest requires a higher generalization level. In the case of behaviors modeling or dynamic events classification, it is advisable to increase the abstraction of the data, designing higher-level descriptions able to model more general structures. In recent years a few interesting works employing machine learning methods showed how these techniques may improve performance in terms of accuracy and efficiency: data-driven approaches are effective to understand possible correlations between measurements and provide systems with the ability of being adaptive. In the field of video surveillance we may exploit the availability of possibly huge sets of examples, acquired by long time observations, and endowed with an internal structure provided by temporal coherence. Within this framework this thesis focuses on: • Studying and developing of robust methods to retrieve space-time information from a video; • Studying and developing of higher-level descriptions, in order to include space-time information within a machine learning framework; • Devising machine learning strategies to model common events and anomalies from huge sets of (possibly) unlabeled examples. These objectives will be addressed both from the algorithmic and the application standpoint and will be integrated in a prototype architecture that combines vision methods for scene perception and analysis and learning techniques for high level description and decision making.