| With the development of society,high-rise buildings and large space sites are constantly emerging in various modern metropolises,but there are also different degrees of fire hazards.Moreover,the fire of large space buildings has the characteristics of fast propagation speed,difficult fire extinguishing and rescue.It is an urgent problem to effectively prevent and detect it as soon as possible.At present,traditional detection methods mostly use sensor detection equipment.Due to many limitations such as poor real-time detection and small detection range,it is difficult to play its advantages in large space buildings.The use of machine vision target detection in the face of complex and changeable,open environment can still accurately and real-time identify the target to be detected.Therefore,based on machine vision,this paper aims to study the early fire smoke in large space building environment.The main research contents include:(1)Firstly,the traditional machine learning smoke detection method based on moving smoke target and multi-feature fusion is adopted.For the input video,a Th-Vibe detection algorithm is proposed to find the motion area,extract the color features,morphological features,background pixel changes and diffusion features when the smoke appears,and use the extracted fusion features to train the support vector machine.Finally,the classification results using non-maximum suppression algorithm to eliminate redundant boxes.(2)Compared with the problems of complex feature design and low detection accuracy in traditional target detection methods,the YOLOv5 deep learning network structure model is analyzed and studied,and a fire smoke detection algorithm based on improved YOLOv5 is designed.The K-means++ algorithm is used to realize the accurate matching between the prior anchor frame and the fire location.Secondly,the attention mechanism channel domain of CBMA is improved,and the backbone network of YOLOv5 is added to improve the feature extraction ability of the network and improve the problem of missed detection of light smoke in the initial fire.The GIo U loss function is replaced by α-Io U loss function as the bounding box regression loss function to improve the accuracy of bounding box location.(3)Finally,the fire detection system software design,according to different fire software interface gives different fire alarm mechanism,and trigger the alarm,with the help of third-party platform to the user mobile phone fire alarm information prompt,in order to take timely action.Through the above design method,the experiment is carried out in the fire video of high and large space environment.The results show that the average precision(AP)values of the YOLOv5 deep learning network model proposed in this paper are 95.2 %,95.8 % and 95.4 % respectively compared with other models,with a maximum increase of 12.5 %.The maximum speed of detection time of single frame fire image on GPU is increased to 27 ms,and the detection indexes are much higher than those of traditional target detection methods.The improved algorithm designed in this study can better achieve real-time,fast and accurate detection and identification of fire in large space building environment.Therefore,the fire identification system designed in this paper meets the practical application requirements.Figure [71],Table [20],Reference [81]... |
PDF Full Text Request