Research And Application Of Mask-wearing Detection Algorithm For Dense Crowds

In recent years,due to the need for epidemic prevention and control,monitoring the proper wearing of masks by pedestrians has been proven to be an effective way to reduce the risk of cross-infection of infectious diseases such as epidemics and influenza in public spaces.However,in crowded places such as train stations,hospitals,and airports,relying solely on staff to oversee mask-wearing is not efficient and puts management personnel at risk of infection.To improve this situation,using computer vision technology to assist management personnel in monitoring mask-wearing can effectively solve these problems.However,the current mask-wearing detection algorithms need improvement,especially in densely populated areas where there are detection difficulties such as target occlusion,small targets,and multi-scale variations in input data.To solve these problems,this study focuses on using the SSD algorithm to detect mask-wearing in densely populated areas.The main work is summarized as follows:(1)Propose an algorithm for mask-wearing detection that improves upon backbone networks,loss functions,and attention mechanisms.Addressing the issue of facial occlusion in densely populated scenarios,where facial features are prone to obstruction,resulting in difficulties in feature extraction and localization.In this paper,a cascaded backbone network is introduced based on the SSD algorithm.VGG-16 and Res Net50 networks are weightedly fused,and a GAM attention module is embedded after each predicted feature layer.This approach enhances the network’s feature extraction capability while paying more attention to extracting critical information from the features,thereby enhancing the model’s perceptual ability for occluded targets.Additionally,the Io U in the localization loss is replaced with Alpha-CIo U to improve the model’s localization accuracy.(2)Present an algorithm for mask-wearing detection that incorporates bidirectional weighted feature fusion and an improved spatial pyramid pooling.In densely populated scenes with numerous detection targets,challenges arise from small-faced individuals and spatial arrangements of multiple targets at varying scales.In this paper,a bidirectional weighted feature fusion(BWFF)mechanism is proposed to address these issues.It combines low-semantic,high-resolution features from shallow layers with high-semantic,low-resolution features from deeper layers to complement each other.Additionally,an improved spatial pyramid pooling technique(RSPPF)is introduced.By incorporating contextual information from local regions and global contexts through different receptive fields,the model can better capture small-scale targets and variations across different scales.Considering the characteristics of small targets,such as limited feature utilization and a small proportion in the dataset,this paper employs various data augmentation techniques to enhance the detection of small targets in mask-wearing scenarios.Techniques such as random expansion and scaling,duplication augmentation,and Mosaic augmentation are utilized to increase the number of small target samples and enrich background information.This approach enhances the probability and diversity of detecting small targets,improving the model’s performance and robustness in various realistic dense scenes for mask-wearing detection.(3)Apply the innovative algorithm model developed in this study to a mask-wearing detection system designed for densely populated areas.The system features a user-friendly interface and convenient,flexible algorithm interface for seamless integration.It provides functionalities for image detection,folder(image)detection,video detection,real-time monitoring,and scene adaptation.For training and evaluation purposes,a total of 4,474 images were selected and re-annotated from the RMFD mask face dataset,an open-source dataset from Wuhan University,and the mask-wearing dataset available on Baidu Paddle Paddle.The dataset was divided into training and testing sets with an 8:2 ratio.In addition to the training and testing data,a subset of data not included in the training or testing sets,along with self-collected videos of crowded subway entrances during peak hours,were used to evaluate the system.The results demonstrate the stability and reliability of the system.Figure [46] Table [9] Reference [75]...