Real-Time Fire and Smoke Detection Using Multimodal Object Detection

Fire and smoke are dangerous to life, damage and the environment and therefore, early detection is necessary to be able to respond to the emergency promptly. The conventional sensor-based fire detectors are known to have delayed response time, low coverage and susceptibility to environmental changes. To overcome these issues, this study suggests a framework of real-time multimodal object detection, which will convert RGB, thermal, and depth data in order to identify fire and smoke. The methodology includes a multimodal feature extraction pipeline, transformer-based cross-modal fusion (CMFF), and an adaptive YOLO detection head, which is predetermined to detect small flames and moving smoke patterns. To minimize the number of false positives, a spatiotemporal consistency module also verifies motion properties between adjacent frames. The experimental outcomes show that the suggested multimodal system attains the mean Average Precision (mAP) of 92.4 that is significantly better compared to the unimodal RGB-only models. Performance of 41 FPS in real time makes it appropriate to be used in surveillance network, edge devices, and smart IoT platforms. These findings provide insights into the multimodal fusion potential to improve the accuracy of early fire detection in different settings.