Pose Estimation Method of Maintainability Test Components Based on PVNet

In order to construct the virtual-real fusion maintainability test scenario and human-machine interaction mechanism, it is necessary to obtain real-time pose information of components. Therefore, an improved PVNet based component pose estimation algorithm was deeply studied, and the improved ResNet18 was used as the main backbone network to obtain the feature map. In the subsequent processing of the feature map, ELU activation function is used instead of LeakyReLU activation function in the original module. Compared with LeakyReLU, ELU is smoother and can handle complex data sets and deep networks. The experimental results show that the improved algorithm achieves a good effect on the self-made Linemod format air filter cover data set, in which the accuracy of 2D mapping index reaches 91.84%, the accuracy of ADD index reaches 86.69%, and the accuracy of 5cm5° reaches 94.73%, effectively improving the accuracy of real-time pose estimation and virtual reconstruction.