Research On Unsafe Behavior Recognition Method Based On Attention And Key Points Of Miners

Coal is one of the main energy sources in our country and is in high demand.Coal accidents occur from time to time.In these accidents,human factors account for a large proportion.Therefore,the identification of unsafe behaviors is of great significance.The thesis uses a sequence of frames to identify unsafe behaviors in the mine.In order to solve the problems of insufficient light,miners' key points are used to improve the algorithm.Meanwhile,this thesis presents a spatial-temporal attention to extract interactive information from miners and scenes.The details of the research are as follows.1)In terms of the representation of underground information,the improved YOLOv3 algorithm is used on our miner dataset to detect miners and environment.The size of a prior box is adjusted to fit the bounding box of miners and identifiers,and the fusion layer is changed to improve the accuracy.In order to solve the problem of missing target,the information is optimized by combining the posture matching algorithm and appearance matching algorithm.Finally,experiments are carried out based on the under-mine dataset,and the m AP is improved by 9.6 compared with the original model.2)In order to improve the influence of lighting conditions on the model,the key point information and RGB information are used to construct effective features.First,the importance of different key points is emphasized by dividing the adjacent matrix,and graph convolution is used to extract coordinate information of key points.Then 3D convolution is used to extract the Gauss heatmap.Finally,the C3 D model is used to identify unsafe behaviors by combining the coordinate and heatmap information.The accuracy of the final model on our dataset is 78.13%,which is 3.13% higher than the initial model.This thesis verifies the effectiveness of key point information in behavior recognition.3)Since miners' unsafe behaviors are easily affected by the environment under the mine,action information is combined with environmental information in the thesis.The C3 D module in fusion model is used to extract information about miners and environment.The forms of initial attention and feature-based attention are proposed in the thesis.It can be seen from the experiment that the accuracy of the attention model is 82.9%,which is 7.9% higher than baseline.Furthermore,the spatiotemporal attention mechanism is proven effective on UCF101 and HMDB51 datasets.The accuracy of initial attention has improved by 1.8% and 5.9% on these two datasets,respectively,and the accuracy of feature-based attention has improved by 2.5% and 4.1%.In summary,the recognition model which is fused with attention and key points has a good effect on the identification of unsafe behavior.The model is effectively improved by combining miners and environmental information.The results show that the algorithm proposed in this thesis has good effect on the identification of unsafe behaviors,and it can reduce the occurrence of accidents under the mine.