Research On Recognition Method Of Unsafe Behavior Of Miners In Mine Belt Area

Coal mining industry is a labor-intensive industry with miners as the main body.The long-tine single and heavy working activities,together with the complex and harsh underground environment lead to great fluctuation of miners' mood and the decline of self-control ability,which causes the occurrence of unsafe behavior of miners.Domestic and foreign scholars have not done research on the recognition of unsafe behavior of miners based on human three-dimensional skeleton points,so it can be a new research direction in the field of miner behavior research.The mine belt conveyor is an important part of the underground transportation system,many miners take risks and carry out some unsafe behaviors in the belt area,such as climbing belt conveyors and relying on belt conveyors,which compose for the main causes of belt safety accidents.Human action recognition technology,based on Kinect sensor,is applied in this article to identify the unsafe behavior of miners in the belt area.After considering the actual environment of the mine and comparing the mainstream behavior recognition technology,kinect is used as a sensing device to capture the miner's somatosensory information and the RGB video of the belt conveyor area.Besides,the coordinate information of 20 skeletal joint points of miners is combined with color images information about belt area to identify unsafe behavior of miners.A representation methodology,based on multi-level complementary features,is proposed to describe miners' dynamic and complex behavior.This representation methodology integrates the relative position characteristics of motion attitude,the angle characteristics of motion attitude and the inter-frame displacement characteristics of motion attitude.In order to reduce the miner's behavioral feature dimension and remove redundant features,this paper removes the features that are less related to miners' behavior.It also uses feature importance evaluation algorithm based on random forest to analyze the standardized behavior characteristics to achieve feature selection.A keyframe sequence search model which depends on cosine similarity to extract keyframes of miner behavior is employed to avoid the relatively large dimension of behavior feature.In order to accurately recognize the miner behavior in the belt conveyor area,two methods,the weighted average voting model and Stacking algorithm in ensemble learning,are introduced.The large-scale belt conveyor dynamic test platform is used to collect the miner behavior data of the belt conveyor area,meanwhile,we demonstrate the validity of the miner behavior recognition model based on ensemble learning on self-built data set of miner behavior.The recognition accuracy of weighted average voting model and Stacking algorithm for seven types of miners' behavior are 91.2% and 92.7% respectively.In order to effectively identify the unsafe behavior of miners,the unsafe behavior judgment model of miners in the belt conveyor area are established.The model is based on the Mask R-CNN convolution network and the Kinect human skeletal joint point coordinate mapping.To further improve the recognition accuracy of miner behavior,a behavior recognition model based on multiple stacking algorithm is proposed.In the model,the recognition accuracy of seven types of miners' behavior reaches 94.8%.Also,the validity and extendiability of the proposed behavior algorithm are verified by UTKinect datasets and Florence datasets.