Real-time fire and smoke detection system for diverse indoor and outdoor industrial environmental conditions using a vision-based transfer learning approach

Uttam U. Deshpande^1*Goh Kah Ong Michael^2*Sowmyashree H. Srinivasaiah³Harshel Malawade¹Yash Kulkarni¹Yash Desai¹

• ¹Department of Electronics and Communication Engineering, KLS Gogte Institue of Technology, Belgaum, India
• ²Center for Image and Vision Computing, COE for Artificial Intelligence, Faculty of Information Science and Technology, Multimedia University, Melaka, Malaysia
• ³Master of Computer Applications Department, Bangalore Institute of Technology, Bengaluru, Karnataka, India

The risk of fires in both indoor and outdoor scenarios is constantly rising around the world. The primary goal of a fire detection system is to minimize financial losses and human casualties by rapidly identifying flames in diverse settings, such as buildings, industrial sites, forests, and rural areas. Traditional fire detection systems that use point sensors have limitations in identifying early ignition and fire spread. Numerous existing computer vision and artificial intelligence-based fire detection techniques have produced good detection rates, but at the expense of excessive false alarms. In this paper, we propose an advanced fire and smoke detection system on the DetectNet_v2 architecture with ResNet-18 as its backbone. The framework uses NVIDIA’s Train-Adapt-Optimize (TAO) transfer learning methods to perform model optimization. We began by curating a custom data set comprising 3,000 real-world and synthetically augmented fire and smoke images to enhance models' generalization across diverse industrial scenarios. To enable deployment on edge devices, the baseline FP32 model is fine-tuned, pruned, and subsequently optimized using Quantization-Aware Training (QAT) to generate an INT8 precision inference model with its size reduced by 12.7%. The proposed system achieved a detection accuracy of 95.6% for fire and 92% for smoke detections, maintaining a mean inference time of 42 ms on RTX GPUs. The comparative analysis revealed that our proposed model outperformed the baseline YOLOv8, SSD MobileNet_v2, and Faster R-CNN models in terms of precision and F1-scores. Performance benchmarks on fire instances such as mAP@0.5 (94.9%), mAP@0.5:0.95 (87.4%), and a low false rate of 3.5% highlight the DetectNet_v2 framework's robustness and superior detection performance. Further validation experiments on NVIDIA Jetson Orin Nano and Xavier NX platforms confirmed their effective real-time inference capabilities, making them suitable for deployment in safety-critical scenarios and enabling human-in-the-loop verification for efficient alert handling.