Robust Workflow Recognition Using Holistic Features and Outlier-Tolerant Fused Hidden Markov Models

Monitoring real world environments such as industrial scenes is a challenging task due to heavy occlusions, resemblance of different processes, frequent illumination changes, etc. We propose a robust framework for recognizing workflows in such complex environments, boasting a threefold contribution: Firstly, we employ a novel holistic scene descriptor to efficiently and robustly model complex scenes, thus bypassing the very challenging tasks of target recognition and tracking. Secondly, we handle the problem of limited visibility and occlusions by exploiting redundancies through the use of merged information from multiple cameras. Finally, we use the multivariate Student-t distribution as the observation likelihood of the employed Hidden Markov Models, in order to further enhance robustness.We evaluate the performance of the examined approaches under real-life visual behavior understanding scenarios and we compare and discuss the obtained results.