Human Abnormal Behavior Recognition Based On Coal Mine Underground Monitoring Video

At present,the monitoring system of coal mine underground control plays a huge role in ensuring the normal comprehensive mining,comprehensive excavation,development and the safety of underground workers.However,the monitoring personnel in the dispatching room watching the monitoring video for a long time will cause visual fatigue,and will miss the inspection of the operation errors and other problems that occur during the operation of coal miners,which will seriously threaten the safety of the operation of underground miners.Although the computer vision technology has been continuously promoted in the field of monitoring video,the complex and changeable terrain and thick dust in the coal mine have also been affecting the detection of abnormal behavior in human posture,and has not achieved ideal results in dealing with abnormal behavior in the mine.At present,scholars from all countries are also studying the methods of human posture estimation and human abnormal behavior detection in coal mines,but the behavior detection of miners in the dark environment of coal mines is a relatively difficult problem.In order to recognize the human posture,the modeling of two-dimensional images to three-dimensional images has been studied,and the accuracy of posture key point extraction and the correlation and fusion between key points have also been studied in depth,In order to recognize the abnormal posture of underground water exploration and drainage in coal mines,a short-term and short-term memory CNN network model integrating attention mechanism is studied,and the feasibility of abnormal behavior detection is verified.The work completed in this paper is as follows:(1)In response to the current research status of low robustness in human pose estimation,a method based on 2D images captured by monocular cameras is proposed,which utilizes depth space information to create conditions for extracting shallow 3D pose models and reduces the ambiguity associated with3 D key points in 2D images in monocular cameras.At the same time,the minimum filtering method is used to denoise 2D pose information,and the key point hierarchical regression module is used to reconstruct the depth of human pose joint point space into 3D space.The correlation between key points is expressed through weight information,and deep fusion is used to extract shallow 3D pose patterns.Finally,the shallow 3D pose model is refined through3 D pose refinement to obtain a complete pose model with refined pose.By conducting experiments on the BMHAD dataset and comparing the PCK values of the generated 2D and 3D pose joint models of CPMs,SH,and CP,it was found that the improved CPMs had the best performance.The overall PCK values of 2D key points reached 95.5%,while the overall PCK values of 3D key points reached 87.2%,demonstrating the effectiveness of the refined pose model in human pose estimation tasks.(2)In view of the detection of abnormal behavior of water exploration and drainage in coal mines,the data set collected during the internship in coal mines adopts convolutional neural network CNN as the basis,integrates long-term and short-term memory networks and attention mechanism,and automatically extracts the human posture features in the video.The experimental results show that the long-term and short-term memory convolutional neural network based on the self attention mechanism can obviously focus on the miners’ body lines,which provides great help for the determination of the miners’ water exploration behavior;In addition,LSMT also has strong ability to describe features and time information,and determine whether behavior is normal or abnormal,demonstrating good performance.