Improved YOLOv5-based Detection For Safety Equipment Wearing And Smoking Behavior In Cement Plant Staff

Production safety is crucial for maintaining the normal production order of a factory.As an important production base for national economy,ensuring safety in cement plants is of utmost importance.However,during the production process,employees occasionally violate rules regarding dress code and smoking,which threatens production safety and employee health.Additionally,the large area and complex environment in cement plants make manual inspection and reminders difficult and inefficient.Therefore,this thesis proposes a method for detecting the safety of employee dress and smoking behavior in cement plants using the improved YOLOv5 algorithm.The research mainly includes the following aspects:1)Safety wear detection based on improved YOLOv5.To solve the problem of overlapping clothing,a new labeling method is utilized to highlight the main features of clothing and avoid the influence of complex environments.In addition,algorithm adjustments are made to adapt to the modified dataset,including adding Coordinate Attention and modifying NMS to Soft-NMS.Coordinate Attention emphasizes the correlation between spatial position information and features to improve the quality and efficiency of feature representation.Soft-NMS adjusts the score of target boxes to better handle problems such as target stacking.The experimental results show that compared with using the original YOLOv5,the improved algorithm has achieved a 4.7% increase in m AP in safety dress detection.2)Smoking detection based on improved YOLOv5.For the detection of slender cigarettes,two aspects are considered to improve detection accuracy: optimizing the label box of the dataset and improving the algorithm.In dataset annotation,the scope of the annotation box is appropriately increased to improve the feature information of the target,thereby enhancing the classification and positioning capabilities of the model.In terms of algorithm improvement,this article uses the decoupled head technology.This technology separates the predicted position and classification of the target.By separating the two tasks,it can reduce miscarriages and errors in locating detection frames,and avoid inaccurate detection frames that in turn induce neural network learning errors.In addition,a method of adding detection heads for small targets was used to further improve the accuracy of smoking detection.The experimental results show that the improved cigarette detection algorithm has achieved a 6.6% increase in m AP compared with the original YOLOv5 algorithm.3)Deployment of models based on Tensor RT acceleration.To address the problem of deploying multiple models for multi-task detection and reduce the computational costs of the improved models,this article uses Nvidia’s Tensor RT to accelerate the proposed algorithms and employs Triton Server to host multiple models.By parallel running multiple models,the inference efficiency of deep learning applications is effectively improved.Actual tests show that the efficiency of the hardware-accelerated algorithm is increased by 68%.The object detection algorithm and network structure optimization designed in this thesis for safety dress and smoking behavior in cement plants can achieve better performance in processing monitoring images.By using GPU parallel computing and other technologies,the inference speed of the model has been accelerated,which can meet the needs of real-time monitoring.It has good algorithm performance and adaptability.