Towards Reliable Edge Vision AI: Addressing Data, Adaptation, and Deployment Constraints

Modern visual recognition systems often fail where reliability matters most, for example, in safety-critical environments characterized by rare events, changing conditions, strict latency, and privacy constraints. This thesis addresses \three fundamental obstacles that limit such systems: Data Scarcity, Domain Drift, and Embedded Deployability. We tackle each obstacle through a dedicated contribution. (i) a Diffusion-based Data Augmentation Framework, (ii) a Video Stream–aware Online Student–Teacher Distillation Paradigm, and (iii) an Edge AI Deployable Stereo Surveillance Solution. First, we propose the Controllable Image Augmentation (CIA) Framework. A Stable Diffusion-based generative pipeline that enriches scarce datasets with semantically faithful, auto-labeled synthetic samples. CIA complements classical augmentations by preserving spatial and semantic integrity, achieving up to a 10% improvement in mAP, and significantly enhancing generalization under data scarcity. Second, we contrast CIA-generated data with physically rendered simulation data from a game engine, revealing a precision–recall complementarity. CIA mixes improve precision through photometric realism, while simulation increases recall via structural diversity. Small (~10%) real-synthetic (generated or simulated) blends yield the best trade-off. Naive Tri-source combinations (real/generated/simulated) introduce bias interference, motivating a curriculum-mixing strategy that progressively introduces data sources based on reliability and domain similarity. Third, to mitigate domain drift under real-world variability, we introduce the Stream-Based Active Distillation (SBAD) paradigm. It's an online, budget-aware learning methodology for auto-adaptive edge deployment. SBAD scores incoming frames by utility, and performs micro-batched updates with temporal coherence verification, and enables continual adaptation under bandwidth and latency constraints. Finally, we present INFRASECURE, an NVIDIA Orin–based Stereo Vision surveillance solution that operates in real time and ensures privacy by design. Instead of transmitting full video, it processes locally, generating event-level alerts and automatic safety reports. Collectively, our contributions materialize within the lifecycle of the INFRASECURE Edge AI system designed for INFRABEL. They establish a coherent data-centric framework that governs every stage of deployment, from addressing data scarcity through controlled augmentation to ensuring real-time adaptability and privacy-preserving operation directly at the edge.