A Large Dataset With a New Framework for Abandoned Object Detection in Complex Scenarios

Abstract

Video surveillance is a crucial part of public safety and security systems. State-of-the-art techniques either use back-tracing for owner identification or there is no provision of reporting ownership of the abandoned objects. We provide a novel object association technique which very elegantly tackles this issue. We propose a convolution neural network based framework for abandoned object localization and owner identification in video surveillance systems that performs exceptionally well on publicly available datasets and our newly developed dataset. For testing the usability and robustness of these systems, an extensive set of test cases is a prerequisite, to this end we provide an elaborate surveillance dataset covering many complex cases which are not available in existing datasets. Also, the developed dataset has been tested with various state-of-the-art techniques, showing the complexity and challenging nature of the cases. Results obtained from the proposed framework are highly motivating as it provides event detection along with the time of abandonment and ownership information, which can help further investigation and prevent unwanted incidents.