Reproduce “IMUNet: Efficient Regression Architecture for IMU Navigation and Positioning” on OpenResearch

About this paper

Behnam Zeinali, Hadi Zandizari, J. Morris Chang

Researchers at the University of South Florida introduced IMUNet, an efficient deep learning architecture for IMU navigation and positioning that processes raw sensor data directly. This model generally surpasses other mobile-optimized CNNs in navigation accuracy across multiple datasets, while consistently demonstrating superior efficiency with the lowest FLOPs, smallest model size (1.4 Mb), and fastest real-device inference latency (387  Sec).

Key insights

• Integrating depth-wise and point-wise convolutions with residual connections in the MobileResNet Block significantly reduces model size and computational cost while maintaining high accuracy for IMU data processing.
• Using the Exponential Linear Unit (ELU) activation function in IMUNet helps mitigate the 'dying ReLU' problem, potentially leading to faster training convergence and improved accuracy in deep IMU regression networks.
• Modern smartphone AR APIs, like Google ARCore, offer a viable and accessible alternative for collecting high-quality ground truth trajectories, facilitating the creation of diverse IMU datasets without specialized hardware.

Summary by alphaXiv.