A Novel Siamese-Based Approach for Scene Change Detection With Applications to Obstructed Routes in Hazardous Environments

Abstract

The demand for automatic scene change detection has massively increased in the last decades due to its importance regarding safety and security issues. Although deep learning techniques have provided significant enhancements in the field, such methods must learn which object belongs to the foreground or background beforehand. In this article, we propose an approach that employs siamese U-Nets to address the task of change detection, such that the model learns to perform semantic segmentation using background reference frames only. Therefore, any object that comes up into the scene defines a change. The experimental results show the robustness of the proposed model over the well-known public dataset CDNet2014. Additionally, we also consider a private dataset called PetrobrasROUTES, which comprises obstruction or abandoned objects in escape routes in hazardous environments. Moreover, the experiments show that the proposed approach is more robust to noise and illumination changes.