| The production environment and mining operations in coal mines have many hazardous factors.To ensure the safety of miners’ lives,it is necessary to constantly monitor their status and movements.Currently,the vast majority of coal mining enterprises are equipped with video surveillance systems.With the rapid development of artificial intelligence technology,the use of intelligent technology in conjunction with manual supervision is of great significance for the safety production and intelligent management of coal mining enterprises.Deep learning models are used in this thesis to detect,track,and recognize whether safety helmets are being worn by miners,restricted areas are being entered,and dangerous moves are being made.The main research content and work are as follows:Firstly,to address the problem of insufficient feature extraction by the model due to low resolution of safety helmets in the view of monitoring cameras,this thesis introduces the bidirectional weighted feature pyramid module and the non-stride convolution SPD-Conv(Space-to-depth Non-Step Convolutional)module into the YOLOv5 s detection model.The baseline model is optimized from the two directions of multi-scale feature fusion and reducing fine-grained information loss.Using a selfmade safety helmet dataset for verification,experimental results show that the helmet detection model with the introduced bidirectional weighted feature pyramid and fused non-stride convolution SPD-Conv modules performs better than other models in terms of accuracy,with m AP reaching 95.3% and 95.8%,and detection speeds reaching 154.9and 156.2 frames per second,respectively.Furthermore,by calculating the intersection over union between safety helmets and human heads,the model can determine whether the safety helmet is being worn,avoiding false detections such as when the safety helmet is held in the hand.Secondly,in response to the problem of missed and false detections of human targets due to similar backgrounds in underground coal mines,this thesis incorporates self-attention mechanisms into the YOLOv5-based detector,using a window sliding self-attention mechanism network to extract pedestrian features.It increases the focus on the target area while extracting global features,thereby reducing the impact of similar backgrounds.By using a self-made underground pedestrian dataset for validation,experimental results show that the improved model achieves an m AP of98.9%,which is a 1.5 percentage point increase compared to the baseline model.The evaluation results on the MOT16 dataset after combining with the tracker show that the pedestrian tracking model designed in this thesis has significantly improved performance.Once again,with regard to the multi-scale changes caused by pedestrians moving in narrow underground passageways from the monitoring perspective,this thesis utilizes a high-resolution feature extraction network based on the YOLOv5s＿swin detection model to extract human skeleton information and accurately extract targets of different scales,and then inputs them into a graph convolutional neural network for the recognition of miners’ actions.Experimental verification is conducted using a self-made action recognition dataset,and the accuracy of four types of actions,namely,removing safety helmets,falling,and others,all exceed 96.4%.And the highest recognition accuracy being for falling,reaching 98.2%.By testing in simulated scenarios,the experimental results show that the model can accurately detect the types of actions performed by workers,which verifies the feasibility of the action recognition model.Finally,a fully functional software for intelligent monitoring of coal mines underground has been designed in this thesis,which combines the three deep learning models mentioned above.By switching video interfaces,safety helmet wearing detection,pedestrian tracking,and action recognition of miners in different scenarios underground are enabled.The thesis has 67 figures,10 tables,and 83 references. |
PDF Full Text Request