BEYOND MULTI-TARGET TRACKING: STATISTICAL PATTERN ANALYSIS OF PEOPLE AND GROUPS

Every day millions and millions of surveillance cameras monitor the world, recording and collecting huge amount of data. The collected data can be extremely useful: from the behavior analysis to prevent unpleasant events, to the analysis of the traffic. However, these valuable data is seldom used, because of the amount of information that the human operator has to manually attend and examine. It would be like looking for a needle in the haystack. The automatic analysis of data is becoming mandatory for extracting summarized high-level information (e.g., John, Sam and Anne are walking together in group at the playground near the station) from the available redundant low-level data (e.g., an image sequence). The main goal of this thesis is to propose solutions and automatic algorithms that perform high-level analysis of a camera-monitored environment. In this way, the data are summarized in a high-level representation for a better understanding. In particular, this work is focused on the analysis of moving people and their collective behaviors. The title of the thesis, beyond multi-target tracking, mirrors the purpose of the work: we will propose methods that have the target tracking as common denominator, and go beyond the standard techniques in order to provide a high-level description of the data. First, we investigate the target tracking problem as it is the basis of all the next work. Target tracking estimates the position of each target in the image and its trajectory over time. We analyze the problem from two complementary perspectives: 1) the engineering point of view, where we deal with problem in order to obtain the best results in terms of accuracy and performance. 2) The neuroscience point of view, where we propose an attentional model for tracking and recognition of objects and people, motivated by theories of the human perceptual system. Second, target tracking is extended to the camera network case, where the goal is to keep a unique identifier for each person in the whole network, i.e., to perform person re-identification. The goal is to recognize individuals in diverse locations over different non-overlapping camera views or also the same camera, considering a large set of candidates. In this context, we propose a pipeline and appearance-based descriptors that enable us to define in a proper way the problem and to reach the-state-of-the-art results. Finally, the higher level of description investigated in this thesis is the analysis (discovery and tracking) of social interaction between people. In particular, we focus on finding small groups of people. We introduce methods that embed notions of social psychology into computer vision algorithms. Then, we extend the detection of social interaction over time, proposing novel probabilistic models that deal with (joint) individual-group tracking.