Skip to main content
Overview
Affiliations
AffiliationTelephone
Associate Professor in the Department of Computer Science+44 (0) 191 33 44556
Co-Director in the Wolfson Research Institute for Health and Wellbeing

Biography

Background

Amir Atapour-Abarghouei is an Associate Professor in the Department of Computer Science at Durham University and Co-Director of the Wolfson Research Institute for Health & Wellbeing, working on dependable computer vision and machine learning that operate robustly outside the lab, with applications spanning biodiversity monitoring, precision agriculture and pollinator health, public health and vector surveillance, trustworthy medical imaging, security/surveillance and sustainable transport. He has led or co-leds AI/vision on major research programmes, including Horizon Europe BioDiMoBot (Autonomous Long-time Aquatic Biodiversity Monitoring Robot, EIC Pathfinder Sensorbees (bio-hybrid environmental sensing), EPSRC IAA 360° video anomaly detection, DSTL persistent wide-area surveillance and an Innovate UK/Evergreen Health Solutions KTP on real-world dermatology AI. He is the Project Lead for the MoniRail KTP (£299,744), which embeds Durham-led AI into MoniRail’s in-service monitoring to predict track and vehicle degradation, optimise condition-based maintenance and enable robust positioning in GNSS-denied settings. 

Methodologically, his group advances efficiency of learning, applications of computer vision and deep learning in robotics, robustness under distribution shift (domain adaptation, multi-task learning, label-efficient training), bias-aware medical AI, and perception in difficult sensing regimes (thermal/infrared, panoramic/360°), with open field-ready pipelines. His work also focuses on semi-supervised anomaly/rare-event detection with his GANomaly/Skip-GANomaly approaches having been adopted in Intel OpenVINO/GETI product line and cited by 40+ patents as the underlying anomaly detection method.  He serves as Associate Editor for IEEE Transactions on Cognitive and Developmental Systems, Area Chair for BMVC and IROS, co-organiser of the CVPR NAS workshop, Chair of the BMVA Computer Vision Summer School at Durham and a member of the BMVA Executive Committee.

Research interests

  • Neuromorphic perception and control: event-based vision, spiking neural networks.
  • Efficient autonomy: modular full-stack pipelines for SLAM, teach-and-repeat navigation, obstacle avoidance, path planning and real-time decision-making on edge/embedded hardware.
  • Robust perception under shift: domain adaptation, label-/data-efficient learning, multi-task learning and temporal consistency for depth/segmentation/optical flow in adverse or unusual sensing regimes (thermal/IR, panoramic/360°).
  • Anomaly & rare-event detection: semi-/self-supervised generative methods, spatial and spatiotemporal anomaly localisation for surveillance, industrial inspection and medical imaging.
  • Multimodal sensor fusion: event+IMU+polarisation and classical vision fusion for reliable state estimation and guidance.
  • Multimodal representation learning: cross-modal alignment and grounding (vision–language–inertial), shared latent spaces and contrastive/self-supervised objectives.
  • Continual & self-supervised learning: Strategies to mitigate catastrophic forgetting and task-agnostic lifelong learning for autonomous platforms.
  • Hardware-aware ML: lightweight NAS, pruning/quantisation, latency/energy-constrained model design for real-time deployment on CPUs/GPUs/neuromorphic chips.
  • Time-series modelling & streaming analytics: multiscale forecasting and change-point detection.
  • Trustworthy ML: bias detection/mitigation, robustness explainability, and safety validation for high-stakes applications (healthcare, transport).
  • Mapping & positioning: event-based VIO/SLAM, map alignment (incl. GNN/diffusion approaches) and GNSS-denied localisation (e.g., tunnels).
  • Scalable training systems: reproducible pipelines, large-scale experimentation, and open field-ready software stacks (e.g., ROS integration).
  • Resource-efficient neural architectures: principled designs for fast stable convergence and high accuracy per flop and spiking/temporal operators tuned for tight memory/latency budgets.

Publications

Chapter in book

Conference Paper

Doctoral Thesis

Journal Article

Supervision students