Research On Workshop Behavioral Safety And Dress Code Detection System Based On Multi-objective Detection

In modern industrial production,the behavioral safety and dress code of shop floor employees are issues that managers pay close attention to.However,traditional workshop safety monitoring measures are usually done manually and cannot achieve comprehensive and real-time monitoring of the workshop production process and make timely warning.As an important preventive measure,the detection of abnormal behavior of workshop employees and the detection of the dress code of safety protective equipment play a crucial role in safe production.In order to achieve real-time detection of abnormal behavior and standardized dress code of workshop employees,this paper takes the behavior and dress code of workshop employees as the detection object,uses deep learning based target detection algorithm and spatio-temporal action detection algorithm for detection,and the main research contents are as follows:(1)A workshop safety helmet wearing detection model with improved YOLOv4-tiny is proposed: SCM-YOLO.Real-time detection is performed for the wearing of common protective equipment such as workshop safety helmets.The Leaky Relu activation function of the original backbone network was replaced by the Mish activation function based on the YOLOv4-tiny model to ensure the information mobility in the feature extraction process.The spatial pyramid pooling structure and convolutional attention mechanism were incorporated to enhance the adaptability of the model to different features and to improve the detection accuracy of small targets.Finally,K-means++ anchor frame clustering,label smoothing,and Mosaic data enhancement algorithms were used in the model training phase to improve the detection speed and generalization of the model.The experimental results show that the SCM-YOLO algorithm improved the m AP by 4.76% compared to the YOLOv4-tiny algorithm on the same safety helmet dataset,and its m AP reached 93.19%.The inference speed reached 22.9 FPS on a GTX 1050 ti general performance graphics card,which can meet the accuracy requirements and real-time requirements of the workshop safety helmet wearing detection task.(2)An efficient hybrid domain Slow Fast network for abnormal behavior detection in the workshop is proposed: DA-Slow Fast.In order to meet the detection of abnormal behavior in the complex environment of the workshop,spatio-temporal action detection is performed for six common categories of abnormal behavior in the workshop: making phone calls,sleeping at the workbench,falling down,running violently,playing with cell phones,and fighting,to obtain information on the category of abnormal behavior as well as to locate the action generators.In this paper,the Slow Fast spatio-temporal action detection model is used as the base model,and the original Slow Fast model is improved for the problems of low accuracy and poor feature extraction of the spatio-temporal action detection algorithm in the complex scenes of the workshop.Firstly,the cross-entropy loss function in the Slow Fast algorithm dealing with action classification task was replaced with Focal Loss to solve the problem of unbalanced various action samples in the video dataset.Secondly,in order to improve the feature extraction and capture fine-grained motion details of the Slow Fast backbone network,the DANet hybrid domain attention mechanism was incorporated in the Slow branch and the two-branch lateral connection to improve the detection accuracy of the network for static actions.Finally,the DA-Slow Fast algorithm achieved a detection speed of 29FPS(Ge Force GTX 2080Ti)FPS on the experimental platform and an m AP of 55.63% on the insecure action dataset,which is 5.17% higher than the m AP of the original Slow Fast model,and has significantly improved the detection capability for static insecure and obscuring behaviors.