Research And Application Of Slab Real-time Detection And Tracking Algorithm Based On Deep Learning

The steel industry is a pillar industry of the national economy,and continuous casting is an important process in steel production.The molten steel after smelting is cast into a slab with a unique number sprayed through a continuous casting mold,and transported to the hot rolling process or intermediate warehouse through a roller table.To ensure that the slab is continuously rolled according to the order order,it is necessary to use the slab information tracking system for monitoring and scheduling.Currently,slab tracking is mainly based on manual or PLC tracking,which requires a large amount of manpower and material resources.This article applies deep learning methods to real-time detection and tracking of continuous casting slabs on site through research on continuous casting slab detection and tracking algorithms,assisting the continuous casting production process system in managing slab information.Firstly,a slab dataset was created for the transportation roller area of the continuous casting workshop,which includes slab targets and crane targets under various working conditions and time conditions.We also introduced the self-made slab dataset,counted the number of images and various targets in the dataset,and analyzed the horizontal and vertical ratios and scales of various targets in the slab dataset.Secondly,to address the issue of poor detection performance of existing universal detection algorithms in continuous casting production sites,an improved YOLOv5 slab detection algorithm based on feature fusion and attention mechanism is proposed.In order to enhance the feature expression ability of the target,CBAM has been added to the backbone network;In order to enhance the ability of cross scale feature fusion,the feature fusion network is replaced with Bi FPN.In order to obtain better network performance,some activation function in the network are replaced with ACON-C functions.In the process of model training,a new regression loss function VCIo U Loss is proposed.The test on VOC dataset shows that the loss function can effectively improve the detection accuracy.The test results on the self built dataset show that the proposed algorithm effectively improves the accuracy of slab detection and achieves better results compared to other detection algorithms.Thirdly,aiming at the existing problems such as complex network,large number of parameters,and inability to deploy effectively at low hardware cost,a lightweight LightYOLOv5 continuous casting slab detection algorithm is proposed.In order to reduce the amount of parameters and computation of the backbone network,Shuffle Net v2 is used to replace the original backbone network.In order to enhance the feature expression ability of the target in complex industrial scenes,the CA attention mechanism is added at the end of the backbone network.To further reduce the parameter amount of the network,DCSP structure is designed for feature fusion network,and a cross-scale feature fusion mechanism DCFPN is designed based on DCSP.The test results on the self-made dataset show that the proposed algorithm can achieve accurate detection of slab object.Compared with the original model,the number of parameters is greatly reduced,and the detection speed of the model is greatly improved,which can meet the real-time and accuracy requirements under the condition of low-cost hardware.Finally,two improved slab detection algorithms and SORT multi-objective tracking algorithm to the tracking of continuous casting slabs are applied in this dissertation.In order to analyze the tracking effect of the proposed slab detection algorithm combined with multiobjective tracking algorithm,on-site video of continuous casting slab to conduct slab tracking tests on the two slab detection algorithms is used in this dissertation.The test results show that the multi-objective tracking algorithm of VAAB-YOLOv5+SORT has higher tracking accuracy,while the multi-objective tracking algorithm of Light-YOLOv5+SORT has smaller computational complexity and faster tracking speed.This dissertation also designed the UI interface and deployed the Tensor RT model to accelerate the tracking speed of the model.The detection and tracking algorithms were applied to the tracking of continuous casting slabs.