Research On Forest Fire Video Recognition Technology Based On Multi-feature And Deep Learning

The National Forest Fire Prevention Plan(2016-2025)refers to the fragility of our forest ecological environment and the urgent need to further improve forest fire prevention capacity,strengthen infrastructure and equipment,improve scientific fire prevention systems,improve overall forest fire prevention and control capacity and minimize the occurrence of forest fires and disaster losses.Therefore,it is of great significance to use efficient methods and techniques to detect and warn forest fires,to detect forest fire videos and to promote the application of computer vision.Starting from the improvement of traditional forest video fire pyrotechnic,this paper carries out the forest video fire recognition research by combining image processing,deep learning and machine learning.The main research work is as follows:(1)According to obtain high efficiency forest video,the algorithm of UAV cruise route planning is studied in order to meet the needs of UAV cruise line in limited time.A IACO algorithm is proposed,GA and PSO are used to improve the heuristic factors of ACO,double pheromone update of ACO is obtained,and target function is formulated to improve the problem that the original iteration falls into the optimal solution slowly.Cruise planning experiments combining UAV cruise constraints showed 4.26% reduction in track costs and9.1% reduction in iterations.(2)A forest fire recognition model combining multi-feature extraction and SVM is studied.In the image segmentation algorithm,the restriction conditions are changed,the YCBCr color model is incorporated in the RGB and HIS color space,and the flame texture and motion salience are determined by the grayscale threshold.The experimental results show that89.9% of forest fires are correctly identified.(3)An algorithm for smoke recognition of forest fires based on the dynamic texture of irregular moving regions is proposed.A training and testing library with dynamic texture of smog was established by selecting white and black smog and videos with distractors in the dataset.The smoke detection is performed by using sliding window,moving perspective is extracted by background subtraction,dynamic texture feature vectors are extracted by CLBP descriptor,and support vector machines are used to train and classify the suspected smoke mass.The experimental results show that the correct recognition rate of frame-level smoke video is 91.8%.Based on this method,the performance of other descriptors is compared from coding mode and collecting points.(4)A multi-feature video target detection network architecture is proposed and designed to mine the features of video in different representation spaces,generate more robust feature diagrams and use them for subsequent target detection tasks.The video is fed into a backbone network based on multi-representation feature extraction to generate a feature map containing complete representation information.The region proposes that the network use the multi-representation feature map to predict potential target areas as constituencies.The target eigenvectors are generated by region maximization based on the multi-representation feature diagram generated by the backbone network and are used to calculate the target type probability and target boundary box regression parameters.Experiments show that the algorithm can also perform well in complex scenarios,and has the ability to quickly judge target attributes to achieve forest fire detection.The accuracy of fire video target recognition is94.1%In summary,this paper combines with many methods of fire and smoke feature extraction to identify forest fire video targets,and innovative methods combine multiple features and deep learning.The multi-feature video detection network improves detection efficiency significantly,reduces false alarm rate and reduces time loss.