Research On Mask Wearing Detection And Tracking Technology For Epidemic Prevention Robots Under Deep Learning Mechanism

At the end of December 2019,a global outbreak of a novel coronavirus pneumonia(COVID-19)occurred,resulting in severe health and economic consequences worldwide.Despite the measures taken by various countries to control the outbreak,the highly contagious nature and multiple variants of COVID-19 make it a difficult problem to solve,requiring constant search for effective solutions to prevent the resurgence of the epidemic.Wearing masks can effectively block the transmission of pathogens through respiratory droplets and serve as a means of bidirectional isolation during the spread of the epidemic.Therefore,in such a social context,this paper utilizes a designed epidemic prevention robot platform to propose a deep learning-based method for mask-wearing detection and tracking specifically for edge devices.The main research contents of this paper are as follows:(1)The current mainstream object detection algorithms were analyzed,and through experiments exploring object detection algorithms,considering the advantages and disadvantages of different algorithms in combination with the requirements of the edge platform in this study,YOLOX-Tiny,which balances real-time performance and accuracy,was selected as the basis for the research.Several improvement measures were proposed specifically for the task of mask-wearing detection.Firstly,the Coordinate Attention(CA)mechanism was introduced into the feature extraction network of YOLOX-Tiny to enhance the extraction ability of challenging and small objects’ detailed information.Secondly,the feature fusion part of the YOLOX network was changed to Slim-neck,which improves the network speed while maintaining accuracy.Lastly,the Varifocal Loss was employed to address the issue of imbalanced sample proportions during model training,thus enhancing the detection performance.The improved mask-wearing algorithm can meet the practical requirements on the edge platform effectively.(2)The paper discusses the SORT and Deep SORT tracking algorithms in multi-object tracking.It introduces the principles of Kalman filtering and the Hungarian algorithm.The Deep SORT algorithm is ultimately chosen as the target tracking algorithm in this paper.In response to issues such as the poor robustness of the Deep SORT algorithm in complex scenes,a lightweight improvement is made to the appearance feature extraction network in Deep SORT,and the GIOU matching strategy is introduced.Finally,in combination with the improved YOLOX-Tiny mask-wearing detection algorithm,the experimental results are analyzed to validate the effectiveness of the algorithm.(3)The deployment of the mask-wearing detection and tracking algorithm on the epidemic prevention robot platform is studied.Firstly,the epidemic prevention robot platform is built.Then,the edge inference framework Tensor RT is used to accelerate the model.Finally,the model is deployed on edge devices for model testing,evaluating the recognition accuracy,processing time,and practical performance of the model.Through testing on embedded platforms,the proposed algorithm in this paper demonstrates good tracking performance for objects with frequent changes and satisfactory detection results for small objects.The algorithm also adapts well to variations in target sizes,exhibiting minimal missed detections and false alarms for small objects.It effectively addresses the challenges of mask-wearing detection and tracking in complex scenarios.