Detecção de comportamento anormal em vídeos de multidão

Abstract

Security systems produce a massive amount of video material that can be used to recog-nize abnormal behavior or activities, which expose the people to life-threatening scenarios. However, human operators are not able to evaluate ali the material available in a consistent manner. So, the automatic recognition of video behavior can be crucial for the effective use of surveillance systems to maintain the security of an area or the integrity of people in pub-lic places. This work presents an method focused on the recognition of abnormal behavior in crowd videos. This method combines feature-based methods with appearance-based methods and use them according to the context of the scene. Appearance-based methods create modela based on the leveis of image intensity, while feature-based methods use data extracted from the image, such as edges, lines and coordinates, to generate their modela. The feature-based approach is generally used because it describes the scene in more de-tails, however it involves higher computational costa. The proposed method displays for the human operator only content with possible crowd agglomeration or dispersing, which are considered abnormal behaviors evaluated in this research. The displayed video receive visual marks to help the human operator to locate suspicious activities identified by the system. The resulta obtained during the experiments show that the proposed method is able to recognize abnormal behaviors in crowd videos and mark areas of the image where abnormalities as agglomeration or dispersion are detected. The proposed method, differ-ent from classical approaches available in the literature, makes assessments of the suspect acenes different from the assessment of the acenes with normal behavior or with only the background. As a consequence, experimental resulta show that the proposed method performs 64% faster than the baselines over a database created for Chis work, as well as 71% faster than baselines on UMN and PETS2009 databases. In addition, the proposed method achieves 90% of accuracy on the YAB database, otherwise the baseline method achieves 85% of accuracy.