Research On Fire Smoke Detection Algorithm Based On Embedded Platform

Fire disaster is one of the main disasters threatening public safety and destroying natural resources.It seriously endangers economic and social development,while threatening people’s safety and health.Smoke is the most significant feature at the beginning of a fire,and accurate and timely detection and identification of smoke is an effective means to prevent fires from causing greater harm.Traditional computer vision algorithms have the problem of low recognition efficiency in smoke detection,and have a high error rate in smoke detection.With the rapid development of deep learning in the field of computer vision,applying deep learning models deployed on embedded platforms to fire and smoke detection is a current research hotspot.However,how to deploy a deep learning framework to an embedded platform and be able to identify smoke using a built smoke detection system outdoors remains a challenge.This is mainly because embedded platforms have limited computing power and can only deploy lightweight deep learning network models.When deploying smoke detection models on embedded platforms,there are problems of slow detection speed and low recognition accuracy.Therefore,this article proposes an improved lightweight YOLOv5 s framework based on YOLOv5 s network architecture,and deploys it to an embedded platform to improve smoke detection and recognition accuracy and reasoning recognition speed.The main research work of this article is as follows:(1)In order to solve the problem that the YOLOv5 s algorithm lacks the ability to extract smoke feature information,this paper proposes a smoke detection algorithm based on spatial attention and channel attention.This algorithm introduces a new attention mechanism in the YOLOv5 s network,improving the algorithm’s awareness of the target area.After the upsampling operation,an optimization mechanism is added to eliminate noise information generated during the upsampling process.Experiments show that with a small increase in the amount of network model parameters,the detection accuracy of smoke images is increased.(2)Aiming at the problems of slow detection speed and low detection accuracy when deploying smoke detection networks on embedded platforms,a lightweight smoke detection algorithm network is proposed.This algorithm enhances semantic information interaction between feature channels by introducing a Ghost module and random grouping operations on feature channels.The algorithm also improves the pyramid pooling(SPP)operation.The improved algorithm reduces the size of the model file by 16.9%,the amount of parameters by 18.3%,and the computational complexity by 20.3% while maintaining the same reasoning accuracy on the smoke dataset.(3)This paper builds a smoke detection system using an embedded platform,which includes modules such as high-definition digital cameras,mobile power supplies,embedded platforms,and external screens.The improved lightweight model is deployed on the embedded platform after being compressed and accelerated by Tensor RT.This system can accurately detect the fire smoke information in the environment.