Detection Of Roadway Tunnel Fires Video With Fusion Of Time-series Features

Road tunnel fire events are unusual events that can seriously affect road traffic safety.To guarantee the normal operation of road tunnels,the algorithm needs to detect fire events quickly and accurately.This is an important reference value for the prevention and control of road tunnel fires.At present,the detection algorithms for road tunnel fires cannot extract the fire characteristics effectively and completely.The detection accuracy of these methods is low and slow.In this paper,we use traffic monitoring video data to research techniques of fusing temporal features in fire video detection based on the extraction of fire image features.The main contents are as follows:1)Construction of road tunnel fire dataset.The study constructed a large-scale highresolution road tunnel dataset classified by scenarios,with the vehicle on fire as the main research object.There are 27 fire scenes in the dataset,including tunnel,highway,road,forest,and homemade fire scenes,with a total of 90,788 data samples.This provides a large amount of data support for the research of fire video detection technology.2)Research on fire detection technique research with temporal feature fusion network.The network can input continuous video sequences.The technique proposes a multi-channel branching approach to extract fire target features from continuous video sequences.To better focus on feature correlation between video frames,the network performs decision fusion of features in the ROI region blocks of the fire targets in the images.This improves the completeness of target feature information and complements each other.The experimental results show that the average accuracy of temporal feature fusion network detection is 84.2%.3)Research on fire detection technique research based on 3D frame difference network.The 3D frame difference network adds a temporal dimension.It processes adjacent frames and their frame difference images with different branches.It uses 2D convolutional structures to extract semantic information of the image and 3D convolutional structures to extract temporal features of the target motion.The two features are later fused.This eliminates the information redundancy between frames,reduces the complexity of the model,and improves the detection performance of the model.The average accuracy of the model on the test set is 91.0%,and the FPS is 63.7.The missing and wronging detection rates of the fire detection analysis strategy in the highway tunnel scene are 2.52% and 2.03%,respectively.4)Research on fire detection technique research with feature selection and aggregation.The study adds a video classification branch at the output of the decoupled YOLOv8-based first-level detector head.Then the output of the primary detector is filtered and fused by feature selection and aggregation network to complete the secondary detection.The experiments show that the average accuracy of the two-level detector is 91.8% and the FPS is 30.5.The average missing detection rate and the average wronging detection rate of the fire detection analysis strategy are 2.28% and 1.13%.The temporal feature fusion technology for fire video detection investigated in this paper shows high detection precision in road tunnel fire dataset and high robustness for different complex scenes.The research can be used to detect abnormal fire events in real road tunnels and other scenes,which can well solve the problems such as difficult fire detection and untimely reporting.