Nikita V. Laptev, Olga M. Gerget, Andrey A. Kravchenko, Vladislav V. Laptev, Dmitry Yu. Kolpashchikov

National Research Tomsk Polytechnic University

The paper presents an algorithm for clustering the areas of predicted areas, implemented in a system for the early detection of forest fires. The system is based on a neural network model for searching for a fire object detection. At the output of the neural network model, an array of bounding boxes, the estimated location of objects of a given class, class labels, and estimates of the probability of belonging to a class are formed. The task of the algorithm considered in the article is to select the desired objects with a bounding box and calculate the average value of the probability of belonging to a class by combining and averaging the bounding boxes that have intersections with the desired object. The article provides a brief overview of existing algorithms for displaying the resulting bounding box on an image. The rationale for choosing a neural network model for an early fire detection system is given. The results of the comparison of the NMS, Soft-NMS algorithms and the area clustering algorithm for solving the problem of detecting a smoke cloud in an image are presented.

object detection, bounding box, algorithm, classification, localization