The DeepILIA research team is pleased to announce that Mohamed Benkedadra has successfully defended his PhD at the Faculty of Engineering (Faculté Polytechnique) of the University of Mons (UMONS) on the 16th of January 2026. His dissertation, titled “Towards Reliable Edge Vision AI: Addressing Data, Adaptation, and Deployment Constraints: A Case Study on Railway Construction Sites,” was conducted under the supervision of Prof. Sidi Ahmed Mahmoudi and Dr. Matei Mancas, within the ILIA and ISIA laboratories at UMONS. He is also a researcher for the Trusted AI Labs initiative.
His research addresses a key challenge in modern computer vision: deploying reliable AI systems in real-world safety-critical environments. While Deep Learning models perform well in controlled settings, they often struggle with limited labeled data, changing conditions, and strict hardware constraints in industrial environments. Mohamed's work focuses specifically on railway construction sites, where AI systems must operate robustly despite environmental variability and operational complexity.
The thesis proposes a data-centric framework built around three main contributions. First, the Controllable Image Augmentation (CIA) framework leverages diffusion-based generative models to generate semantically consistent synthetic training data and alleviate data scarcity. The framework can be used here : https://github.com/multitel-ai/CIA
Second, the Stream-Based Active Distillation (SBAD) paradigm enables edge-deployed models to continuously adapt to changing environments through selective learning from incoming video streams. Here's the library : https://github.com/fennecinspace/SBAD
Finally, the research introduces INFRASECURE, a real-time edge AI surveillance system running on NVIDIA Jetson hardware that performs on-site perception and generates safety alerts while preserving privacy by processing video locally. Check it out here : https://www.youtube.com/watch?v=K_eQTDkdTew
The thesis was evaluated by an international jury composed of Prof. Thierry Dutoit (UMONS) as president, Prof. Sidi Ahmed Mahmoudi (UMONS) as supervisor, Dr. Matei Mancas (UMONS) as co-supervisor, and jury members Prof. Mohammed Benjelloun (UMONS), Prof. Patrick Le Callet (Polytech Nantes), and Prof. Sergio Rodriguez (Université Paris-Saclay).
Conducted in collaboration with INFRABEL and supported by the TRAIL / ARIAC initiative, the work bridges academic research and industrial application. By combining generative AI, continual learning, and edge computing, the thesis contributes to the development of reliable and deployable vision systems for real-world safety monitoring and infrastructure applications.