Research And Application Of Fire Smoke Detection Algorithm Based On Deep Learning

The traditional fire image detection algorithms mostly use flame as the detection target,and the smoke is generated at the beginning of the fire,and the flame will be generated in the middle of the fire,which makes it impossible to prevent and control the first time.At the same time,the common algorithm for identifying smoke with a single feature decreases the recognition rate when the scene is complex or has a lot of interference,and the missed detection rate increases sharply.In view of the above research status and application scenarios,this paper proposes a smoke detection algorithm based on the fusion of moving targets and multiple features without CUDA and cuDNN.First,the input video is used to detect the motion area in the video by using the inter-frame difference and background difference detection algorithm.The smoke is extracted as the color characteristics of white smoke,black smoke,blue smoke,gray smoke,and the smoke is often convex due to the wind Morphological features,dynamic features of smoke and texture features,and then use the extracted fusion features to train a support vector machine.The extracted motion area is used to classify smoke and non-smoke in turn through the classifier.Finally,the non-maximum suppression algorithm is used to remove the redundant frames based on the classification results.The test videos in the forest environment,the rural environment and the urban environment obtained 92.44%,90.62%,and 91.14%accuracy,respectively,which is a significant improvement over the traditional algorithm where the accuracy is less than 85%,and the hardware computing power requirements are low.It can be transplanted in Raspberry Pi,FPGA,DSP and other hardware environments.This paper also proposes a forest fire smoke detection algorithm based on KCF and YOLOv3.Under the condition of supporting CUDA and cuDNN,the dual-thread detection method is used.On the one hand,the DarkNet deep learning framework is built,and the YOLOv3 target detection network is built.The depth model is trained by calibrating the smoke and flame in a large number of forest fire sample images.On the other hand,build a KCF tracking thread.After thread 1 detects flame or smoke,it sends the detected position information and category information to thread 2,and then uses thread 2 to perform target position detection on the target position information detected by thread 1,calculates the thread response and finds the response value The largest detection frame is used as the target frame to obtain its confidence information.Finally,it is compared with the detection result of thread one.It overcomes the bad influence of the algorithm by illumination and deformation,and improves the accuracy,robustness and adaptability of the target tracking algorithm.In this paper,the method of model compression and model reduction is used to simplify the YOLOv3 detection model.The detection speed reaches 56FPS in the 8700K+Titan X environment,and 15FPS on the embedded development board Nvidia Jetson Nano.Real-time performance is achieved after frame skip detection The practicality of the algorithm is verified.