An Improved Magnetic Tracking Approach Based on ResNet-LM Fusion Algorithm
Pengfei Guo,Houde Dai,Qianhui Yang,Qiaoyuan Huang,Hanchen Yao
DOI: https://doi.org/10.1109/tim.2022.3216597
IF: 5.6
2022-11-11
IEEE Transactions on Instrumentation and Measurement
Abstract:Permanent magnet-based tracking (PMT) approach is a reliable solution for motion tracking and navigation. The conventional PMT for tracking one or more magnets is based on optimization algorithms such as Levenberg–Marquardt (LM) algorithm. However, the tracking accuracy and computation time of the LM algorithm depends on the initial values of pose parameters. The artificial neural network (ANN) based methods can provide an alternative solution for optimization algorithms, whereas the studies based on machine learning algorithms had limited tracking performance. In this study, we proposed a ResNet-LM algorithm, based on the fusion of the deep learning algorithm and optimization algorithm, to improve PMT performance. First, we employ a residual neural network (ResNet-20) to track a single magnet. The tracking accuracy of ResNet is ( mm, ) by padding with a scale of 2, where the size of the magnetic flux density (XT) becomes (3, 8, 8). Second, the prediction result of ResNet is adopted as the initial value for the LM algorithm, with the improved tracking accuracy of ( mm, ). Finally, we utilize an adaptive weighted fusion algorithm to fuse the prediction results of ResNet and LM algorithms. Experimental results show that the pose tracking accuracy of ResNet-LM is improved to ( mm, ), where the mean computer time for the ResNet-LM algorithm was 79.3 ms. The ResNet-LM-based pose accuracy for AGV Parki- g was ( mm, ). The proposed approach can improve the tracking accuracy and reduce the computational time of PMT, thereby extending the application scenarios of PMT.
engineering, electrical & electronic,instruments & instrumentation