Research On Video Fire Detection Method Based On Two-Stream Convolutional Neural Network

Fire detection is of great significance in ensuring the safety of people and property.If the flame can be detected in the early stage of flame combustion,the damage caused by fire and explosion can be greatly reduced.Early sensor-based fire detection methods have the disadvantages of single information,small detection range,and misjudgment.With the rapid development of computer vision technology,vision-based fire detection technology can overcome the limitations of sensor-based fire detection technology and effectively improve the reliability of fire detection.The visual fire detection methods can be divided into fire detection methods based on image processing and fire detection methods based on deep learning.Fire detection methods based on image processing require a complex manual feature extraction process and have great limitations for different scenarios.The convolutional neural network can take the original image as an input and automatically learn features from the image,but the convolutional neural network with only the image as the input ignores the flame motion information.Based on the two-stream convolution neural network,the video fire detection method studied in this thesis is oriented to the fire prone scene,such as factory,shopping mall,forest and so on.The main contents include the following parts:(1)In order to use optical flow to express the motion characteristics of the flame,an improved optical flow estimation algorithm is proposed to overcome the shortcomings of the traditional optical flow algorithm in the calculation error of illumination,noise and large displacement movement.First,redefine the data items of the optical flow model by combining the brightness constant hypothesis and the Hessian matrix constant hypothesis,introducing a non-square penalty function and bilateral filtering,then using the dual algorithm to greatly reduce the number of iterations in the calculation process,and finally using multi-resolution The layered and refined method makes the algorithm adapt to the situation of large displacement movement.A comparative experimental study was conducted on the test data set,and the results show that the method in this thesis can better maintain the image edges,and is more robust to illumination changes and image noise,and the overall performance of the algorithm is better.(2)In order to improve the real-time performance of the algorithm,it is necessary to filter out most non-flame areas in the video,and a method for extracting suspected flame areas based on motion features and color features is proposed.First,the Vibe +algorithm is used to detect the moving target to extract the moving area,and then the moving area is subjected to color feature detection combining RGB and HSI color models.The area that meets the color detection conditions is determined to be a suspected flame area.The above method reduces the time complexity of the algorithm and ensures the real-time performance of the algorithm.(3)Aiming at the problem that the traditional deep learning-based fire detection method does not combine flame motion information,a fire detection method using twostream convolutional neural network is proposed.First collect flame images and videos as the experimental data set;then use two-stream convolutional neural network to combine the spatial and temporal features of the video to classify and identify the suspected flame regions,and finally output the flame detection region as the final detection result.A comparative experiment on the test data set verifies the effectiveness of the proposed method.(4)Based on the above theoretical methods,a fire detection software prototype system was designed and developed using Py Qt5,and the proposed methods were integrated.The main modules of the software system include a video sequence loading module,a suspected flame area extraction module,and a fire detection module.The fire detection software prototype system was used to extract the suspected flame area and fire detection of the video in the real scene,which further proved the effectiveness of the method proposed in this thesis.The video fire detection method based on the two-stream convolutional neural network proposed in this thesis has achieved considerable results in the fire detection task.It can contain the fire in the initial stage,and is of great significance in ensuring personal and property safety.