Obstacle Detection Of Underground Rail Transpotation In Tungsten Mine Based On Deep Learning

China possesses abundant tungsten resources,making it the world’s leading country in terms of tungsten reserves.The exploitation of tungsten is crucial for meeting the global demand in the tungsten market.Transportation within tungsten mines is a critical step in the mining process.Currently,most underground tungsten mines employ rail transportation systems,which are susceptible to geological deposits and falling rocks.Miners often traverse these tracks or tunnels to reach their work areas,resulting in potential obstacles and an increased risk of accidents involving tramcars.Therefore,the effective detection,timely warning,and appropriate handling of track obstacles in underground mines are essential for enhancing the safety of mine track transportation and optimizing intelligent transportation systems in smart mines.This paper presents an overview of obstacle detection technologies for track areas in both domestic and international contexts.Currently,there is a limited availability of open datasets specifically designed for obstacle detection in mine track transportation,posing a challenge for further research in this field.Most existing studies on obstacle detection focus on surface scenes under favorable lighting conditions,typically involving large tracks such as trains and subways,where environmental factors have minimal interference.However,the detection of obstacles in mine track transportation remains relatively underexplored.To address these issues,this paper employs deep learning techniques to conduct an in-depth study on underground mine track obstacle detection,specifically focusing on a tungsten mine in Jiangxi Province.The main objectives of this research include:Due to the scarcity of publicly available image datasets for mine track obstacles,this paper addresses the issue by conducting research and creating a tungsten mine underground track obstacle dataset based on real-world conditions in tungsten mines.The study begins by examining the unique environment of mine rail transportation and carefully selecting appropriate equipment for data collection.Images of the rail transportation area are captured using the underground rail system of a tungsten mine in Jiangxi Province as the background.Considering the challenges posed by uneven illumination and strong noise interference in underground tungsten mines,this paper simulates real working conditions by introducing noise into the captured images.Various grayscale processing methods and filtering techniques are compared to determine the most suitable image preprocessing approach for this study.Data expansion techniques are employed to increase the dataset size,and Z-score normalization is applied to ensure that the obtained image data adheres to a normal distribution.Finally,the expanded and processed image data is manually labeled using the Label Img software,resulting in the creation of a comprehensive dataset required for the present study.In the context of mine track obstacle detection algorithms,this paper conducts a comparative analysis of five algorithms,namely SSD,YOLOV3,YOLOV4,YOLOV5,and YOLOX,with the aim of identifying the most suitable algorithm for mine track obstacle detection.The theoretical aspects of these algorithms,including the network structure,feature extraction design,level of feature detection,as well as their respective advantages and disadvantages,are analyzed and compared.Subsequently,experimental verification is performed by implementing the algorithms in a Py Torch-based framework and training and testing them using the custom dataset created for this study.The results are thoroughly analyzed and discussed,while the model’s performance indicators are assessed and evaluated.Through a comprehensive consideration of theoretical design,training,testing,and model evaluation,it is demonstrated that the 2D target detection algorithms can be successfully applied to mine rail transportation obstacle detection.Among the studied algorithms,YOLOX is identified as particularly suitable for detecting obstacles in underground tungsten mines.This paper introduces the application of the monocular 3D detection algorithm in the detection of mine track obstacles.Considering the unique characteristics of the mine track environment and the limitations of the monocular 3D detection algorithm,this study focuses on improving the monocular FCOS3 D detection algorithm and proposes a novel approach called the multi-branch-Moni3 D algorithm.To enhance the performance of the algorithm,several modifications are made.Firstly,the Res Net module in the original network is replaced with an Efficient Net module incorporating the SE(Squeeze-and-Excitation)module.Additionally,the FPN(Feature Pyramid Network)structure is replaced with Bi FPN(Bi-directional Feature Pyramid Network).Secondly,the classification process in the head prediction is replaced with direct prediction,reducing the time required for positive sample identification.The proposed algorithm is evaluated using the Nu Scenes dataset,and the results demonstrate notable improvements.Specifically,compared to the original algorithm,the inference time on a single NVIDIA RTX 3080 graphics card is reduced from 79 ms to 27 ms.Although there is a slight decrease in mean Average Precision(m AP)by 1.71%,the monocular 3D detection algorithm shows an increase in mean Average Translation Error(m ATE)by 2.57%,mean Average Scale Error(m ASE)by 4.4%,Mean Absolute Error(MAE)by 6.72%,mean Average Vertex Error(m AVE)by 11.1%,and mean Average Orientation Error(m AOE)by 27.5%.