Fire Object Detection and Tracking Based on Deep Learning Model and Kalman Filter

Automatic fire detection is an interesting challenge for several researches particularly in video surveillance application. Nowadays, it is required to exactly locate and recognize fire regions to avoid damage as soon as possible and to overcome the available fire detection methods limitations. Indeed, it is proposed in this work to combine the advanced computer vision techniques with the deep learning models. The presented approach aims to detect and track fire regions in video sequences. A fire detector is firstly developed by tuning the tiny-You Only Look Once version 3 (YOLOv3) network hyperparameters. The detected fire regions identified by their bounding boxes positions are then fed into an accurate fire tracker to avoid the missed detections caused by the detector. It is based on exploiting the Kalman filter to relocate the fire area with a more accurate estimation of its position. The overall performance of the novel approach, evaluated on the created dataset, is clearly enhanced from 94.70 to 96.70% in terms of true positives rate. This is thanks to the association of the fire tracker with the detector. This effective combination is demonstrated in experimental results where the number of the recognized fire objects is increased with a reduction of the false positives to only 0.0155 compared to the fire objects number detected without the tracker.