dc.description.abstract
Fire-safety assessment and firefighter training require rapid interpretation of hazards under uncertainty, degraded visibility, time pressure, and high cognitive load. Recent advances in artificial intelligence (AI) and computer vision (CV) enable automated detection of visual hazard indicators such as fire, smoke, extinguishers, and alarm systems. However, such outputs are often presented as labels, bounding boxes, or confidence scores without sufficient contextual explanation. In safety-critical environments, this limits trust calibration, interpretability, and the ability to assess whether an AI output should inform human judgement. Mixed reality (MR) offers the possibility of presenting AI-generated information directly in the user's field of view. This thesis therefore investigates how AI-based hazard detection can be translated into in-situ MR guidance through local visual explanations, i.e. explainable AI (XAI).Following a Design Science Research (DSR) approach, the thesis developed and evaluated a prototype instantiation of an MR-XAI translation-layer framework. The prototype consisted of a YOLO-based hazard detection pipeline, a FastAPI backend for inference and explanation generation, and a Unity-based MR client implemented on the Meta Quest 3. The system supported interactive prototype-level hazard detection, MR-based visual overlays, and on-demand explanation generation using CAM-based attribution methods, including EigenCAM and Grad-CAM++. The prototype was evaluated in two formative case-study rounds with 16 participants. The thesis-focused analysis compared unaided hazard identification with an XAI-aided MR condition. Quantitative data included task completion time, simplified NASA-TLX-based workload scores, perceived situational awareness, and trust/reliability ratings; qualitative participant feedback and researcher observations were used to interpret usability and technical limitations.The results show that integrating AI-based hazard detection and local visual explanations into an MR interface is technically feasible at prototype level. However, the XAI-aided condition did not improve task completion time or average perceived situational awareness compared with unaided inspection. Across 13 valid paired timing observations, unaided inspection was consistently faster than the XAI-aided condition. Simplified NASA-TLX-based workload scores remained in the low-to-moderate range, but were higher for the XAI-aided condition than for the unaided baseline in the second round. Qualitative feedback nevertheless indicated that explanations could support interpretation and understanding when detections and explanations were correct, stable, and spatially readable.A central finding is that the usefulness of explainability in MR depends on the reliability and stability of the complete perception-explanation pipeline. Participants trusted explanations more strongly than the underlying hazard detections, while perceived system reliability was lower. Missed detections, misclassifications, disappearing overlays, delayed explanations, and interaction ambiguities reduced trust and limited practical usefulness. Explainability alone therefore cannot compensate for unreliable perception or unstable interface behaviour. This thesis contributes a working MR-XAI research prototype, formative evaluation knowledge on workload, perceived situational awareness, trust, and interaction challenges, and design implications for explainable MR systems in safety-critical environments. For the firefighting domain, the work highlights the potential of in-situ explanations to support hazard interpretation, while showing that robust detection, stable spatial presentation, low interaction burden, and clear uncertainty communication are prerequisites for operational relevance.
en