Design Of Embedded Video Smoke Detection System Based On Feature Fusion

The frequent fire not only has a great impact on the natural environment,but also causes threat to life and loss of property.Since smoke appears firstly in the smoldering phase at the beginning of a fire,detecting smoke quickly and accurately is the key to preventing fire and reducing loss.Compared with traditional smoke sensing detection methods,video-based smoke detection technology has the advantages of short response time,low cost,and large coverage area,which has attracted extensive attention from researchers in recent years.In order to meet the real-time and effectiveness requirements of the video smoke detection algorithm in the embedded platform,this paper proposes a suspected smoke area extraction algorithm based on the combination of Vi Be foreground detection and color feature discrimination.Firstly,the video frame sequence is preprocessed,and then based on the motion characteristics of the smoke,the Vi Be foreground detection algorithm,which has a small amount of calculation and a more complete detection area,is used to initially extract the moving target area in the video.Then our algorithm carries out morphological operations to further eliminate holes and noise of the smoke image.After that,the distribution characteristics of each component of the smoke image in the YUV space are counted and analyzed,and the color feature judgment criteria of the smoke are proposed,and then the smoke area is screened out from the moving target area.Finally,a 24*24 square image block is used to scan the entire video frame pixel by pixel to block the suspected smoke area,and then obtain the suspected smoke image block for subsequent smoke feature extraction and embedded machine learning classification.This paper designs and implements a set of embedded video smoke detection system based on support vector machine.First,based on the suspected smoke region extraction algorithm proposed in this paper,58 kinds of Uniform Pattern LBP(Local Binary Pattern)features are extracted from the suspected smoke image blocks to form feature vectors,and then sent to the support vector machine for classification,and finally the algorithm obtained the results of smoke detection by transplanting it to the embedded platform “Industri Pi”.Through the detection of smoke test videos in different scenes,the results show that the algorithm based on the support vector machine can obtain an accuracy rate of 89% in video of the near scene,and the accuracy rate of more than 80%in video of the distant scene,and the detection speed is 15 FPS,which verifies the effectiveness and real-time performance of this algorithm.Based on the suspected smoke region extraction algorithm proposed in this paper,this paper also uses feature fusion based on convolutional neural network and image structure similarity to complete the design of embedded video smoke detection algorithm,and transplant it to Industri Pi system to achieve real-time performance.First,based on the Le Net-5 network,the smoke detection convolutional neural network structure of this article is established,and the extracted suspected smoke image blocks are sent to the convolutional neural network for classification.Then analyzed and compared the nature of the image structure similarity between the smoke image and the non-smoke image and its background model,and found that the value of the structure similarity of the smoke image decreases with the appearance of smoke in the initial stage,and with the smoke thicker,and the value tends to be within a certain range and greater than 0.75,and then a smoke discrimination criterion based on the similarity of image structure is proposed.Finally,the structure similarity criterion is used to further judge whether it is smoke after the classification of the convolutional neural network.In the Industri Pi.By comparing the smoke detection effects in different videos,the results show that the embedded video smoke detection algorithm based on feature fusion has a detection accuracy of more than90% in the near scene,and the detection in the distant scene is also generally higher than the support vector.The machine learning algorithm embodies the robustness and effectiveness in this paper.