THE SMOKE DETECTION FOR EARLY FIRE-ALARMING SYSTEM BASED ON VIDEO PROCESSING AND CNN

Abstract

In recent years, large-scale wildfires have occurred one after another around the world, which reminds us of vehicle spontaneous combustion accidents that are closely related to our lives. with the development of new energy vehicles, the number of electric vehicles has increased year by year. However, due to many uncertain factors, the phenomenon of spontaneous combustion of electric vehicles occurs always, especially when electric vehicles parked in the outdoor parking lot spontaneously ignite, when a fire is formed, it is difficult to find out in time, once the fire spreads, it is difficult to extinguish it in a short time. The control of fire mainly lies in prevention. If some existing technology is used to identify and issue an early warning when the fire occurs, it can be extinguished before the fire expands. Smoke is the most obvious feature of a flame before it burns, so smoke detection can play a role in preventing fires from being unburned. Smoke has semi-transparent characteristics，its shape and texture characteristics are easily changed by external interference. These characteristics determine the difficulty of smoke detection. This thesis mainly uses computer vision and CNN technology, to detect and recognize the smoke generated by the early fire in the parking lot. The main research contents of this thesis are as follows. 1. Through the collection and sorting in the preliminary preparation stage, 12 sections of smoke videos used to detect the effect of the algorithm, include 12,470 smoke images and 12,902 non-smoke images used to training and test the CNN. 2. Analyze the general characteristics of the smoke in the early stage of the fire from the static (color) and dynamic aspects of the fire smoke. 3. Use CNN as a classifier to recognize smoke images. The work of this thesis has the following innovation: I. Data set. The data sets used in this thesis are the video clips recorded in the outdoor environment. The data sets can simulate the scene of the real fire smoke in the real situation, and it is more practical. II. Smoke color feature extraction. In this thesis, I focus on the characteristics of HSV color space in smoke, which avoids the feature failure caused by the low quality of video in RGB color space. III. CNN is used to detection smoke. Neural network can automatically learn the characteristics of the images through large number of training data, to avoid the trouble of the artificial selection of a single feature. IV. Finally, a complete set of outdoor fire smoke detection system is designed, the system can be achieved on the outdoor of fire smoke early warning. Then, four groups of contrast experiments are used to verify the effectiveness of the proposed algorithm. The experimental results show that the proposed method has higher accuracy rate, lower false alarm rate and missing detection rate than other methods, also proves that the application of CNN to the field has certain research value. At the end of this thesis, the advantages and disadvantages of the proposed algorithm are summarized, the corresponding solutions and the direction of further research are put forward according to the shortcomings of the algorithm.