Research On Personnel Detection And Statistical Analysis In Coal Mine Complex Scenes Based On Deep Learning

The coal industry is the mainstay industry of the national economy,mineral production safety is also a matter of great concern to the state and society.Manually viewing surveillance videos takes time and effort,it is no longer sufficient to meet actual production needs.Therefore,Coal mine surveillance videos intelligent analysis technology has become the one of important research contents of Intelligent Mine.Personnel detection and personnel counting are important research topics of coal mine surveillance videos intelligent analysis field.Real-time personnel detection can ensure the safety of personnel,avoid entering dangerous areas,and ensure the safety of coal mine production.Meanwhile,real-time personnel statistics can meet the daily management of personnel,and it is also of great significance for emergency rescue.Coal mine auxiliary shaft is an important channel for the lifting of personnel and materials.This thesis uses two surveillance videos at the upper and underground wellheads of the coal mine auxiliary shaft.The main work of this thesis is as follows,(1)Building object detection models combining with lightweight feature extraction network for coal mine auxiliary shaftAccording to the requirements of the complex scene of the coal mine,this thesis builds object detection models combining with lightweight feature extraction network to achieve real-time personnel detection.Firstly,this thesis makes two datasets based on surveillance videos,image preprocessing is also necessary.Then,we train two personnel detection models for two different scenes.People are crowed at the underground wellhead and the color of people is similar to the background,detecting people's head replaces of detecting people's entire body.When training the model,in order to reduce the number of iterations and avoid overfitting,uses the pre-training model and combines my own dataset for the second training.Afterwards,we change the width multiplier to implement model compression and compare the impact of model compression on personnel detection.Finally,Experiments show that compared with the original SSD model,the MobileNetV2_SSD model has the same detection precision,and the parameters are only one-fifth of the original,and the detection speed is more than twice that of the original model.(2)Achieving the personnel counting method based on multi-frame detection results for coal mine auxiliary shaftTo achieve personnel counting at the upper and underground wellheads of auxiliary shaft,this thesis proposes a counting method which sets a smaller counting area and contrasts five frame detecting results.The inter-frame personnel correlation solution uses the jaccard overlap between the two neighboring sequence images and the centroid is calculated to determine the moving direction.For accurate counting,the counting area is set on the way that people enter the cage.By testing single batch of people from surveillance videos in two scenes,we verify the accuracy of the counting method.Experiments show the accuracy of the upper wellheads is 95.2% and the accuracy of the underground wellheads is 94.2%.(3)Counting the amount of personnel falling and personnel lifting within 24 hours and drawing the real-time curve of the number of underground personnelDue to coal mine safety regulations,people can only enter the cage from one side and leave the cage from the other side at the upper and underground wellheads of auxiliary shaft.Detecting and counting people in the surveillance videos of people entering the cage at the upper wellheads,we get the amount of personnel falling;Detecting and counting people in the surveillance videos of people entering the cage at the underground wellheads,we get the amount of personnel lifting.Starting from the same time,the amount of personnel falling and personnel lifting within 24 hours will be counted.At the same time point,the number of the two will be subtracted,and the real-time curve of the number of underground personnel will be obtained.Experiments show the accuracies of both scenes are over 94%,the number of people underground are also basically in line with the actual situation.This thesis achieves personnel detection,personnel statistical analysis and the analysis of people underground in coal mine scenes,which has certain practical significance for coal mine safety production.