The Network Output Background Subtraction (NOBS) Algorithm for Unattended Luggage Detection

Abstract

In recent years, a number of algorithms for unattended luggage detection have been proposed. They use both classical computer vision and neural network approaches. Classical methods do not work well on crowded scenes and when long time tracking is required. Neural networks face the fact that unattended objects are very diverse, that requires large representative training dataset and heavy network architecture. In this paper, we propose the Network Output Background Subtraction (NOBS): a real-time algorithm based on the use of a lightweight neural network trained on general purpose dataset with no use of any specialized dataset of abandoned luggage. It allows you to use the advantages of the neural network approach (the ability to track objects more accurately and for a long time) and partially get rid of its disadvantages: the need for an extensive training data set that reflects the large variability of detected objects, and high resource requirements.