Cities are working to make the lives of their citizens as safe as possible. This desire is accompanied by the provision of new devices such as video surveillance cameras in quantity. However, too many of these devices can be counterproductive, as too much information can make it difficult to identify important information from the totality of information available.

The objective of this thesis is first to find solutions allowing to indicate concretely the type of abnormal situation taking place in order to focus the operator's attention on this situation. This part is based on static analyses extracted from video streams of video surveillance cameras. The second objective of this thesis aims at improving the first one thanks to an individualized follow-up allowing to keep the anonymity of the followed persons. This solution is based on displacement prediction using Kalman filters. Finally, the last objective of this thesis is directly linked to the lack of information available to test the efficiency of our different methods. This objective is based on the implementation of a simulation system allowing to get closer to a real situation by using autonomous virtual entities.