BM-UNet: lightweight multiclass-fire segmentation network and robust methods of aiming to the fire based on state-machine of mean segmentation mask and contour center-mass

Abstract

The article presents the BM-UNet neural network architecture optimized for mobile devices. The article focuses on the possibility of real-time operation, takes into account various techniques for lightening the model, and provides a comparison with UNet-half. An improved model in terms of performance/accuracy ratio is applied in the algorithm of frame-by-frame segmentation of the flame on video, the result of which is averaged, and the optimal extinguishing point is found. For the latter, an approach to organizing a finite state machine is presented for switching between time-averaging windows for the possibility of timely response and minimization of mechanical losses of targeting. Combined, the method of frame-by-frame segmentation of flame contours and the algorithm for finding the center-mass point can be used in robotic flame extinguishing systems.