Robust Visual Behavior Recognition

In this article, we propose a novel framework for robust visual behavior understanding, capable of achieving high recognition rates in demanding real-life environments and in almost real time. Our approach is based on the utilization of holistic visual behavior understanding methods, which perform modeling directly at the pixel level. This way, we eliminate the world representation layer that can be a significant source of errors for the modeling algorithms. Our proposed system is based on the utilization of information from multiple cameras, aiming to alleviate the effects of occlusions and other similar artifacts, which are rather common in real-life installations. To effectively exploit the acquired information for the purpose of real-time activity recognition, appropriate methodologies for modeling of sequential data stemming from multiple sources are examined. Moreover, we explore the efficacy of the additional application of semisupervised learning methodologies, in an effort to reduce the cost of model training in a completely supervised fashion. The performance of the examined approaches is thoroughly evaluated under real-life visual behavior understanding scenarios, and the obtained results are compared and discussed.