Técnicas baseadas em similaridade para análise visual de videos de segurança

Abstract

Surveillance camera systems based on CCTV (closed-circuit television) are widely employed in a variety of society segments, from private and public security to crowd monitoring and terrorist attack prevention, generating a large volume of surveillance videos. The manual analysis of these videos is unfeasible due to the excessive amount of data to be analyzed, the associated subjectivity, and the presence of noise that can cause distraction and compromise the comprehension of relevant events, impairing an effective analysis. Automatic summarization techniques are usually employed to facilitate this analysis, providing additional information that may guide the security agent in this decision making. However, these strategies provide little/no user interaction, limiting his/her comprehension regarding the involved phenomena. Furthermore, such techniques only address specific scenarios, in the sense that no approach is good for all situations. In this sense, it is important to insert the user in the analysis process, as they provide the additional knowledge to effectively perform the events identification and exploration. Visual analytics techniques represent a potential tool for such analysis, providing video representations that clearly communicate their content, potentially revealing patterns that may represent events of interest. These representations can significantly increase the capacity of the security agent to identify important events, and filtering/exploring those that represent potential alert situations. In this project we propose a methodology for visual analysis of surveillance videos that employs Information Visualization techniques for events exploration. We specifically coordinate point-placement techniques and Temporal Self-similarity Maps (TSSMs) to create an analysis environment that reveal both structural and temporal aspects related to event occurrence. Users are able to interact with these layouts, in order to change the visualization perspective, focus on specific portions of the video, among other tasks. We present experiments in several surveillance scenarios that demonstrate the ability of the proposed methodology in providing an effective events summarization, the exploration of both the structure of each event and the relationship among them, as well as their temporal properties. The main contribution of this work is a surveillance visual analysis system which provides a deep exploration of different aspects present on surveillance videos regarding events occurrence, providing an effective analysis and a rapid decision making.