Research On Multi-target Detection And Tracking Algorithm For Downhole Miners

Our nation is a major producer of coal resources,and in the process of transforming our energy structure towards a more environmentally-friendly approach,coal will remain the mainstay of our country’s energy supply.As domestic coal mining conditions become increasingly complex,there has been a growing emphasis on mine safety work.Therefore,the research on miner multi-target detection and tracking algorithm based on underground surveillance video is of paramount importance in improving the accuracy and efficiency of coal mine safety operations.Due to the high levels of dust,high humidity,poor lighting,and complex backgrounds present in downhole environments,traditional multi-target detection and tracking algorithms struggle to extract target features effectively.Therefore,many scholars and researchers have applied deep learning-based multi-target detection and tracking algorithms to underground video surveillance.To address the aforementioned issues,this paper proposes the joint improvement of the YOLOv5s-GAD and Deep SORT multi-target tracking algorithms for underground miner tracking.The research content of this paper is as follows:(1)The dim lighting,high humidity,and high dust content present in coal mines make data collection and labeling more challenging.As a result,existing public datasets such as PASCAL VOC and MS COCO are inadequate for meeting the unique needs of underground coal mining environments.This paper utilizes a custom underground miner dataset,Miner21,and enhances its images through a custom saturation transformation model-based dehazing algorithm to mitigate issues such as dim lighting and excessive noise in video images captured by cameras.Additionally,to address the color masking caused by yellow and fine dust in images,the brightness component is enhanced using white balancing and atmospheric light in blurred images.To ensure the validity of the experiments despite the small size of the dataset,this paper employs various image enhancement methods,including random cropping,adaptive image scaling and stretching,and Mosaic data augmentation,to expand the dataset samples.(2)This paper proposes a multi-scale fusion object detection algorithm,YOLOv5s-GAD,based on the improved YOLOv5 s.The proposed algorithm is evaluated on the preprocessed Miner21 dataset.Firstly,four attention mechanism models,STN,SENet,ECA-Net,and CBAM,are compared.Then,ablation experiments are conducted on Ghost Conv,DWConv,and ECA-Net to verify the effectiveness of the fusion of various modules.Additionally,comparative experiments demonstrate that YOLOv5s-GAD achieves an object detection accuracy(AP@0.5)of 98.2% and an FPS of 140.2 frames/s.Furthermore,the model parameters are significantly reduced,and both training speed and accuracy are markedly improved.Finally,through visualizing the heat map,the positions of underground miners focused on during feature extraction are clearly displayed,which confirms the effectiveness of the proposed algorithm.(3)In response to the issue of personnel obstruction during collaborative action of miners,this article proposes the replacement of the shallow residual network in the pedestrian re-identification part of the original Deep SORT algorithm with the full-scale network OSNet,which facilitates omnidirectional feature learning for re-identification purposes.Through comparative experiments,the improved Deep SORT algorithm has demonstrated an enhancement in tracking accuracy by 2.9%and a decrease in the number of identity transitions(IDs)by 1161,while the frame rate(FPS)has increased by 12 frames per second.Finally,a two-point boundary discrimination experiment is conducted to address the problem of tracking miners crossing boundaries in hazardous underground areas,which shows a certain level of auxiliary effect on avoiding such situations.This paper provides a new approach for the multi-object detection and tracking of miners in underground video surveillance.Experimental results have shown that the proposed algorithm can meet the real-time and accuracy requirements of multi-object tracking of miners in underground mines.The research findings can provide a reference basis for improving the safety of coal mine operations and the safety of miner behavior.Figure [29] Table [7] Reference [85]...