Edge-AI-Powered Hazard Detection: A Real-Time Approach for Identifying Obstructions in Emergency Evacuations : Leveraging Monocular Vision for Efficient and Automated Emergency Evacuation Systems

Sammanfattning

Context: Effective emergency evacuation in indoor environments hinges on the rapid identification of obstructed exits to safely guide occupants and first responders. Although computer vision and edge computing have advanced, existing IoT-based solutions rarely address real-time detection of partial or complete blockages in evacuation routes. Objective: This thesis seeks to (1) survey computer vision approaches within the IoT domain for obstruction detection, (2) design a vision-based solution capable of detecting blockage in emergency passways and exits, and (3) evaluate the real-world effectiveness of an IoT-edge integrated prototype for real-time obstacle detection and alert generation. Methods: A semi-systematic literature review following PRISMA guidelines identified a gap in exit-blockage research and highlighted YOLO as the leading framework for real-time, edge-compatible object detection. Guided by Design Science Research and structured using the CRISP-DM process, we manually curated and augmented a bespoke dataset of 4,299 fire and smoke images and 4,742 door images. Two YOLOv11 variants (nano and small) were trained and benchmarked across multiple model formats (PyTorch, ONNX, OpenVINO, NCNN). Performance metrics included mean average precision (mAP50-95), frames per second (FPS), and false-alarm rate. Results: The YOLOv11 nano model achieved real-time inference at approximately ~3.6 FPS on a Raspberry Pi 5, while maintaining competitive detection accuracy (mAP ≈ 60%) and manageable false-alarm rates. Benchmarked model formats revealed trade-offs between throughput and precision, with OpenVINO and NCNN offering faster inference at a slight accuracy cost. Live video tests exposed challenges with intense light sources and occluded doors, informing adaptive IoU thresholds and multi-sensor fusion strategies. Conclusion: This work delivers a systematic methodology for IoT-edge computer vision in evacuation contexts, a publicly documented dataset and benchmarking framework, and critical insights into model trade-offs and failure modes. Future research should broaden dataset diversity, explore advanced overlap metrics, and integrate complementary sensors to enhance robustness in real-world deployments.