Research And Application Of Security Detection Algorithm For Interior Elevator Scenes

Public safety and fire protection are essential components in the construction of a smart city.Once a fire breaks out in closed places such as buildings,it will cause irreversible losses to the safety of people’s lives and property.Therefore,fire safety monitoring between community buildings is particularly necessary.With the rapid development of deep learning,artificial intelligence technology has made significant breakthroughs.How to combine artificial intelligence technology with smart cities and apply them to urban construction has become a hot topic of current research.This topic is a school-enterprise cooperation topic.We have conducted a close cooperation with Party A’s company,combined with actual projects,and proposed solutions for fire safety prevention using the staircase scenario as the application scenario.This scheme improves and tests existing detection algorithms and strategies,which can more effectively monitor behaviors that endanger fire safety,it is of great significance for ensuring fire safety represented by stairways.The main research content of this article is as follows:(1)The YOLACT and CycleGAN based dataset construction and preprocessing methods are used to expand the data of the self-built dataset.Firstly,the ResNet module of the YOLACT instance segmentation algorithm is optimized,and the RFB module is incorporated into the model.At the same time,a more effective Swish activation function is used to improve the nonlinear expression ability and feature extraction ability of the model,improving the accuracy of image data sample segmentation.Then,the segmented human or object target samples are randomly fitted to the background samples to build an extended dataset.Secondly,the CycleGAN style transfer learning method is used to optimize the loss function,making the generated human hair color,clothing style,and battery car data samples closer to actual data samples,enhancing the robustness and generalization of the detection model;(2)An improved smoking detection algorithm based on YOLOv5 is proposed for closed smoking and fire prohibition places represented by elevator lobbies.By adding a small target detection layer,incorporating a convolutional attention mechanism module into the backbone network,and optimizing the loss function,the model pays more attention to small target characteristics.The mosaic data enhancement method used in the YOLOv5s input is improved to a 16-graph random mosaic data enhancement method,which enhances the robustness and generalization of the model against small and dense targets.In response to the increased amount of parameters in the model,the traditional convolution method is replaced by deeply separable convolution.The channel is pruned to make the model more lightweight and improve the detection speed.The experimental results show that the recall rate and average accuracy(IoU=0.5)of the improved YOLOv5s detection algorithm are 4.94%and 6.9%higher than the original YOLOv5s algorithm on the self-built dataset,respectively,and the detection frame rate is 17.1 FPS higher.The improved YOLOv5s detection algorithm has better recognition ability for small targets represented by cigarettes in indoor scenes such as stairways;(3)A new adaptive knowledge distillation strategy for complex samples and a lighter YOLO model has been proposed to solve the problem of electric bicycles entering the room,represented by elevator room scenarios.The experimental results show that this method is superior to traditional distillation methods,enabling the student network model to learn the knowledge of teacher network models,and may even exceed the detection performance of teacher network.This method is suitable for general lightweight network models and significantly improves the feature learning ability of student networks using lightweight networks.After testing,the trained student network model is deployed to the edge computing end.The precision performance is not only better than the baseline network,but also better than the student network refined by the traditional distillation algorithm.The speed performance is far superior to the teacher network,providing a new feasible scheme for the engineering application algorithm.