Approach state detection of lightweight steel sections based on improved YOLOv8-pose

Aiming at the problems of poor real-time and low productivity of the manual detection method for the proximity state detection of the section and the sizing baffle, a lightweight section proximity state detection method based on the improved YOLOv8-pose is proposed. First, GhostConv is used to replace part of the Conv in the Backbone of the original YOLOv8-pose, and GSConv is used to replace the Conv of Neck, and the improved RepGhostBottleneck is used to design a new C2f_RepGhost module to replace the original C2f module to realize the lightweight model, Reduce the number of parameters of the model and easy to deploy. Secondly, to address the problem of target scale change, the ContextAggregation attention mechanism is introduced, and the global context information is effectively integrated and utilized. Third, the original SPPF is replaced by the improved GGP_SPPF so that the model can better understand the overall structure of the image, combine global and local information and reduce the amount of computation. Finally, the original loss function is replaced with the MPDIoU loss function to optimize the regression problem with dataset imbalance and bounding box. The experimental results show that the improved model improves the mAP by $0.6 \%$ compared with the original model, and the number of parameters of the model, the amount of computation, and the size of the model are reduced by $42.54 \%, 34.94 \%$, and $37.5 \%$, respectively, in the homemade steel section and fixed-edge baffle approaching state dataset. Comprehensively, it shows that the improved model improves the detection accuracy while significantly reducing the number of parameters, computation amount and model size, and meets the requirements of edge computing devices, which can effectively find out the required key point information on the steel section and the fixed-edge baffle plate, simulate the edge information when the two are close to each other, so as to judge the proximity state of the two, and compute the pixel distance between the two when they are not close to each other, and judge whether it is in the maximum permissible error range has practical application value.