Xavier BOU HERNANDEZ
Change Detection in Video and Image Time Series
Abstract
This thesis addresses change detection in video and remote sensing image time series — a longstanding challenge in image processing with key applications in surveillance, natural disaster response and land cover monitoring, among others. The diversity of data types and use cases has led to a wide range of methodological approaches. We begin by proposing a taxonomy of change detection, structured around the data source — video or remote sensing — and the nature of changes — blind or semantic-aware.
The first part of the thesis focuses on traditional blind change detection methods, which exploit temporal resolution to model the scene’s background. We review the core principles of such approaches through the seminal ViBe algorithm. Then, we address the problem of reducing the false alarms in video-based blind change detection methods, and later adapt the ViBe mechanism to remote sensing time series for flood detection using Sentinel-1 SAR image sequences.
The second part of the thesis covers semantic-aware change detection. The first two contributions focus on specific problems of object detection, which serves as a foundation for temporal change analysis. First, we explore the use of robust visual and vision-language features for few-shot object detection in high-resolution satellite imagery. Moreover, we propose an orientation representation for oriented object detection that is robust to discontinuities caused by angular periodicity. Lastly, end-to-end semantic change detection is explored by using additional unlabeled acquisitions on top of single-date semantic segmentation datasets.
Key Words
Video ; Deep learning ; modelling
Supervision
Jury
- Sébastien LEFÈVRE, Professor, Université Bretagne Sud (Reviewer & Examiner)
- Marc VAN DROOGENBROECK, Professor, University of Liège (Reviewer & Examiner)
- Loïc LANDRIEU, Research Scientist, École des Ponts ParisTech (Examiner)
- Pascal MONASSE, Research Director, École des Ponts ParisTech (Examiner)
- Silvia VALERO, Associate Professor, CESBIO (Examiner)