Research On Recognition Of Abnormal Behavior In Underground Mine Based On Machine Vision

Coal is the main energy source in our country.The safe production of coal mines is an important factor in ensuring the normal supply of coal.Most coal mine safety accidents are related to the abnormal action of the operators.Therefore,starting from the action of the operators,reducing the occurrence of safety accidents is to improve the safety of coal mines.At present,the coal mine video surveillance system still uses the traditional RGB camera combined with manual supervision,which cannot identify abnormal action in a timely and effective manner,and there are delays and omissions.In order to solve the above problems,this paper proposes a stacking action recognition method based on skeleton information and combined with an improved underground image enhancement algorithm to build an abnormal underground action recognition system based on Jetson Xavier embedded equipment.main tasks as follows:First,according to the characteristics of the human skeleton,a geometric constraint model is proposed to punish the illegal length or angle in the estimation result;for the problem of loss of spatial information in the downsampling of the neural network,HRNet is used to maintain high-resolution representation throughout the entire network training process.And adopt a multi-resolution fusion method to extract more feature information;and use the grouping normalization method to optimize the network,improve the feature extraction ability of the network in small batches,and achieve more accurate pose estimation results.Experiments show that the average joint error estimated on the Human3.6M data set is 41.4mm,which is about 4mm higher than the estimation accuracy of the reference network.Then,aiming at the weight distribution problem of the primary learner in the Stacking integrated learning algorithm,a Stacking action recognition algorithm based on the choice of the learner is proposed.The weight of the prediction result of the primary learner when input into the meta-learner is determined by the confidence of the primary learner,and the confidence is determined by the accuracy of the primary learner's prediction,that is,the primary learner with better performance has a higher weight.The recognition accuracy of 10 action on the Human3.6M data set is 92.3%,which is 2.4% higher than that of the traditional Stacking algorithm,which proves the effectiveness of the algorithm.Finally,aiming at the problem of poor image quality in underground mines,a degraded image enhancement method based on Retinex theory and improved guided filtering is proposed.The low-frequency component estimated after wavelet decomposition and bilateral filtering is used as the estimated value of the illuminance component to achieve brightness enhancement,the edge detection result is used to improve the guide image and the multi-scale decomposition and reconstruction are combined to achieve detail enhancement;based on Jetson Xavier Hardware,based on the above algorithm,implemented an abnormal action recognition system in underground mines.Experiments were completed on self-built action data sets and real underground mine scenes.The action recognition accuracy rate reached 90.5%,indicating that the system can effectively identify the underground mine scenes.Abnormal action improves the safety production level of coal mines.The paper has 55 pictures,9 tables,and 91 references.