End-to-End CNN-Based System for Human Detection in Fire Scenes Deployed via a Flask Web Application

28 May
Posted on Author

Authors: Dr. Rajkumar, Md Nasir Hussain, Harsh Kumar, Shashikesh Kumar, Aman Mehra, Shubham Mahto, Irah Khan

Abstract: Fire emergencies continue to claim thousands of lives and cause enormous economic damage every year across residential, commercial, and industrial settings. While conventional fire alarm systems based on smoke or heat sensors remain the dominant approach in modern buildings, they are inherently reactive, slow to respond in large spaces, and entirely blind to the presence and location of human occupants inside hazardous areas. This paper presents a complete, deployment-ready fire and human detection system built around two complementary deep learning models: a custom binary Convolutional Neural Network (CNN) trained to classify fire and non-fire scenes, and the YOLOv8 Nano one-stage object detector configured to localise human occupants in real time. The outputs of both models are fused inside a Decision Engine that generates one of four contextual threat levels — SAFE, PERSON DETECTED, FIRE ONLY, or HIGH RISK — along with rich visual annotations including bounding boxes, hazard overlays, and severity banners. The system is delivered through a Flask REST backend hosted on Hugging Face Spaces and a React-Vite single-page application that streams live webcam frames entirely through the browser using the HTML5 getUserMedia and Canvas APIs, thereby removing the need for server-side camera access. On the evaluation set, the CNN achieves 96.38 % accuracy, 97.20 % precision, 95.50 % recall, and an F1-score of 96.34 %, while the YOLOv8 Nano detector attains a mean Average Precision at IoU threshold 0.50 (mAP50) of 89.2 %. Frame-skipping and lazy model initialisation raise real-time throughput from 5 FPS to 25 FPS and cut CPU utilisation from 98 % to 35 % on a standard cloud CPU. The results demonstrate that a carefully engineered lightweight architecture, even without GPU acceleration, can serve as a practical, scalable, and cost-effective intelligent safety monitoring solution.

DOI: https://doi.org/10.5281/zenodo.20426948

</p