Enhanced Surveillance: Triple Background Subtraction with YOLO V8

Abstract

In congested areas like malls, airports, train stations, etc., video surveillance facilitates monitoring and provides a sense of security. There is a need for advancements in video surveillance technology to be more robust and efficient. Due to increasing terrorist and criminal activities, addressing the unattended static artefacts on public premises has become a high-priority task. To mitigate human and financial loss, abandoned objects should be dealt with the utmost priority. Identifying abandoned or removed objects in surveillance footage proves challenging due to its complexity, driven by occlusion and sudden alterations in lighting.  This paper proposes a novel technique for detecting and classifying abandoned objects, particularly bags. The work aims to automatically detect abandoned objects. The method involves a robust triple background subtraction technique that extracts background using three sub-models. A Convolutional Neural Network (CNN) -based classifier is used to classify abandoned artefacts. You Only Look Once YOLO V8 is used as the classification algorithm. After the foreground is extracted, graph-based segmentation is used to extract candidate static objects.  Final static objects are extracted using the stability rank calculation method. The suggested approach is validated on three benchmark datasets: PET 2006, PET 2007, and i-LIDS AVSS. Performance parameters include precision, recall, and accuracy. In realistic environments and factual situations like poor illumination and occlusion, the proposed solution outperforms the existing methods. The proposed methods help in the reduction of false positives, reducing the false alarm rate.  The proposed method reaches an accuracy of 99.5%, precision of 93%, and recall of 90%, much higher than earlier proposed systems.