Design Of Cabin Fire Detection System Based On Improved YOLO Algorithm

About 70% of the ship fire accidents happen in the engine room of ships.The internal environment of the engine room is complex,and there are a large number of flammable and explosive substances.Also,the external passages of the cabin are narrow.Once a fire occurs,it will spread rapidly and is extremely difficult to put out,seriously affecting the safety of crew members’ lives and property.Therefore,it is of great significance to study a method to detect smoke and flame in the cabin as early and accurately as possible for ship safety.The video fire detection algorithm based on deep learning has the advantages of a large monitoring range,high detection accuracy,not restricted by space and distance environment,and easy to save data.However,the existing fire detection algorithms are mostly verified in some proprietary data sets,which are not suited to ship engine rooms.Applying such data sets directly in the case of the engine room may cause low recognition and poor real-time performance.This thesis focuses on the research of video image-based flame and smoke detection algorithm for real-time fire warning in ship’s cabin.The proposed algorithms are deployed in the embedded software to complete the test.The main research contents and results are summarized as follows:A special dataset for ship cabin fire recognition is constructed,and a data enhancement algorithm is proposed.There are few fire images about ship cabins in the existing public fire dataset.Therefore,the fire images which are matched to the cabin background are screened out from the cabin fire videos and public fire datasets by analyzing the characteristics of cabin images.After labeling the figures,the cabin-specific dataset Ship Fire Data is constructed,with more than 23,260 images in total.The Growup data enhancement algorithm is proposed to expand the dataset,improve the accuracy of the model by 1.1%,and enhance the recognition capability of the model for small targets.Cabin-specific fire detection algorithms SF-YOLO and GSF-YOLO are designed.The feature extraction network SF-Net is constructed based on the CSPDarknet53 network by introducing a self-attention mechanism.And this is used as Backbone to propose a ship cabin fire detection algorithm SF-YOLO using a multi-scale prediction framework in YOLOv4.In order to reduce the computational cost and the power consumption of the ship,a lightweight model GSF-YOLO is obtained by replacing the normal convolutional module with the Ghost module in the backbone network.The miniaturized network reduces the number of parameters and computational power by 57% and 64%,respectively,compared with the original network.SF-YOLO and GSF-YOLO are trained using intersection over union-based loss function and transfer learning.The accuracy of SF-YOLO and GSFYOLO models are 83.1% and 79.8% respectively with the input resolution of 83.1% and79.8%,and the inference speed is 27.7 frames per second and 41.7 frames per second respectively on NVIDIA GTX 1660 Ti.The model is proved to have strong anti-interference ability and robustness by detecting complex scenes with obscured flames and blurred smoke.Under the condition of balancing accuracy and speed,the ship fire detection algorithm proposed in this thesis is better than Rpi Fire and other algorithms.Considering the power consumption and cost on ships,the Tensor RT tool was used to further compress the proposed networks,doubling the inference speed.The proposed networks are then integrated and deployed into a cabin fire detection software made by QT and tested with fire and non-fire videos in various cabin scenarios.The results show that the detection algorithm designed in this thesis can rapidly and accurately identify fires in ship cabins.