Research On Track Detection Model Of Underground Roadway Scene Based On Deep Neural Network

At present,China’s underground roadway transportation is mainly driven by manually driven rail locomotives.Frequent safety accidents have brought huge economic losses and bad social impact.The automation and unmanned transportation of underground transportation will become an inevitable trend for future development.The realization of computer vision detection of track is an indispensable basic supporting technology.The traditional track detection technology relies heavily on the characteristics of manual design,the detection accuracy and speed are difficult to satisfy,and the robustness is poor.Aiming at the above problems,the thesis carried out research on underground track detection based on deep learning theory.A track detection algorithm based on deep neural network is proposed.The main research contents are as follows:(1)Carry out research on related theories of convolutional neural networks,generative confrontation networks,semantic segmentation,knowledge distillation,etc.,to provide strong support for the design and implementation of subsequent network models.At the same time,the underground track video was collected and preprocessed and annotated,and the underground track training data set was constructed.(2)A downhole track detection algorithm combined with a multi-scale conditional generation adversarial network is proposed.This algorithm addresses the deficiencies of the existing convolutional neural network(CNN)based algorithm used in track detection and the drawbacks of traditional hand-made feature algorithms,Combined with the advantages of generating confrontation networks in image generation,a track detection algorithm is designed.The algorithm uses secondary structure in the generator to ensure that the generator can finally generate high-resolution images.At the same time,the use of multi-scale structure in the discriminator,combined with multi-task learning strategy based on parameter sharing.Finally,Monte Carlo search was introduced to improve the effect of image generation.Experimental results show that the detection accuracy of this algorithm in downhole track reaches 94.27%,which has the advantages of accuracy and strong robustness compared with mainstream detection algorithms.(3)An underground track detection algorithm based on attention distillation is proposed to further meet the real-time and applicability requirements of track detection in actual scenarios.The algorithm uses the superiority of SCNN(Spatial Convolutional Neural Networks)in the lane line detection algorithm,so it is used as a teacher network;ENET(Efficient Neural Network)as a semantic segmentation network,its model is small and fast,so it is used as a student network perform Attention Distillation(AD).Finally,the ENET-AD track detection algorithm is proposed.Algorithm In the detection of lane line data sets and underground track line data sets,ENET-AD has demonstrated its excellent performance,and the test frame rate can reach 41.8FPS,and the detection accuracy rate can reach 91.59%.In summary,the two track line detection algorithms proposed in this thesis can effectively solve the problem of complex track line detection in underground roadway scenes,which lays the foundation for the subsequent realization of advanced underground auxiliary driving and fully automatic unmanned driving technology.