Efficient Semantic-Based Vehicle Retrieval in Long-term Car Park Videos

The proliferation of video data has resulted in huge potential in Big Data technology and applications for video surveillance, security, multimedia tools and analytics. However, large-scale video data over a long period of time necessitates an efficient representation such that the task of retrieving video shots is rapid and accurate. This paper proposes an efficient and comprehensive framework for semantic based vehicle retrieval from long-term car park videos. Colour and motion semantics are respectively retrieved using intuitive colour term and sketch-based trajectory querying. The contribution of this work is twofold. First, we present a strategy for extracting the dominant colour for similarity matching against the Munroe ground truth tuples. Secondly, our proposed framework introduces a unique sketch-based method of retrieving vehicle motions, which relies on user-drawn trajectories. Using the spatio-temporal atom representation for extraction from videos, our approach obtained reasonably good precision scores at very fast retrieval speeds based on one-month long of daytime car park videos.