Preintegrated IMU Features For Efficient Deep Inertial Odometry.

摘要

MEMS Inertial Measurement Units (IMUs) are inexpensive and effective sensors that provide proprioceptive motion measurements for many robots and consumer devices. However, their noise characteristics and manufacturing imperfections lead to complex ramifications in classical fusion pipelines. While deep learning models provide the required flexibility to model these complexities from data, they have higher computation and memory requirements, making them impractical choices for low-power and embedded applications. This paper attempts to address the mentioned conflict by proposing a computationally, efficient inertial representation for deep inertial odometry. Replacing the raw IMU data in deep Inertial models, preintegrated features improves the model's efficiency. The effectiveness of this method has been demonstrated for the task of pedestrian inertial odometry, and its efficiency has been shown through its embedded implementation on a microcontroller with restricted resources.