Automatic video categorization for massively large corpora: A paradigm shift for applications in lane tracking.

摘要

With the rapid development of file sharing and mobile access devices and systems, data generation and distribution has become so convenient and accessible for individuals that users are overwhelmed with information. Research in driver behavior modeling and transportation vehicle systems is migrating in this direction with an increasing amount of data available to the research community. Researchers are urged to reflect on shifting the current research paradigm from focusing on well organized, small amount and high quality data sets in order to face the challenge of emerging extensive, realistic and diverse data set. In this study, we consider an alternative direction which represents our understanding of this new research paradigm. As an application in a specific area, we propose an automatic video categorization framework for the application in lane tracking as preparation for large-scale data analysis. The proposed system provides a fast and effective strategy to screen vehicle videos and categorize segments of video into three-state categories measured by specific features according to video characterization and frame consistency. In order to evaluate the performance of the automatic video categorization system, evaluations are performed using two popular lane tracking systems based on video segments into a three-state space partition. The evaluation shows that the proposed automatic video categorization system is able to separate video segments into quality categories for the specific lane tracking task according to user defined video segment length, and the required amount of data in a selected quality category. Finally, the proposed system will also be released to the public via the web as open source for research purposes on the UTD-CRSS website.