Deep Learning for Estimating the Mounting Angle of Inertial Sensors in Mobile Devices

Core Concepts

A deep learning model is proposed to estimate the yaw mounting angle of an inertial measurement unit (IMU) sensor fixed to a mobile device inside a moving vehicle, using only the IMU measurements as input.

Abstract

The authors present a data-driven approach using deep neural networks (DNNs) to estimate the yaw mounting angle of an inertial measurement unit (IMU) sensor fixed to a mobile device inside a moving vehicle. The key highlights are: A dataset is created by collecting over 52 hours of driving data with an external IMU sensor mounted at a known yaw angle. To expand the dataset, the recorded IMU samples are synthetically rotated to simulate a wide range of mounting angles. A convolutional neural network architecture called MountNet is designed and trained in a supervised manner to regress the yaw mounting angle from the IMU measurements. The model is trained using a custom loss function that accounts for the periodic nature of angles. An algorithm is proposed to smooth the noisy output of the DNN model and enable real-time deployment. The smoothed output converges to an accurate estimate of the mounting angle within 30 seconds of driving. Experiments are conducted to validate the performance of the proposed approach on a validation dataset with synthetic rotations, as well as on a test dataset using a physical smartphone device. The model is shown to estimate the mounting angle with a mean absolute error of 4 degrees within 27 seconds. A comparison is made with an existing off-the-shelf solution that fuses IMU and GNSS data. The proposed DNN-based approach is shown to outperform the existing solution in terms of convergence time and accuracy, without relying on GNSS inputs.

Stats

The dataset consists of over 52 hours of driving data collected with an external IMU sensor mounted at a known yaw angle. The training dataset contains 672,858 samples, and the validation dataset contains 102,023 samples.

Quotes

"Accurate estimation of the device mounting angle is required to rotate the inertial measurement from the sensor frame to the moving platform frame to standardize measurements and improve the performance of the target task." "Finding an IMUs mounting angle has also been studied outside the realm of position and navigation. The authors of [23] present an approach to estimate the mounting of pipeline inspection gauges." "To the authors' best knowledge, this is the first work to suggest an end-to-end solution based on DNNs to find the yaw mounting angle of an IMU sensor."

Key Insights Distilled From

Deep Learning for Inertial Sensor Alignment

by Maxim Freydi... at arxiv.org 04-11-2024

https://arxiv.org/pdf/2212.11120.pdf

Deeper Inquiries

How could the proposed approach be extended to estimate the full 3D orientation (roll, pitch, and yaw) of the inertial sensor relative to the vehicle frame

To extend the proposed approach to estimate the full 3D orientation (roll, pitch, and yaw) of the inertial sensor relative to the vehicle frame, the deep learning model can be modified to output three angles instead of just the yaw angle. This would involve training the model with ground truth data that includes the full 3D orientation information. The dataset creation process would need to be adjusted to include synthetic rotations not only for the yaw angle but also for roll and pitch angles. The model architecture would need to be updated to handle the additional output dimensions, possibly requiring a more complex neural network structure to capture the relationships between the sensor measurements and the three orientation angles.

What other sensor modalities, in addition to IMU, could be leveraged to improve the accuracy and robustness of the mounting angle estimation

In addition to the IMU data, other sensor modalities could be leveraged to improve the accuracy and robustness of the mounting angle estimation. One possible sensor modality is a magnetometer, which can provide information about the Earth's magnetic field and help in determining the orientation of the sensor relative to the magnetic North. By fusing data from the magnetometer with the IMU readings, the model can better estimate the mounting angle, especially in scenarios where the vehicle is stationary or moving at a constant velocity without significant acceleration changes. Another sensor modality that could be beneficial is a camera or vision sensor. Visual odometry techniques can be used to track visual features in the environment and estimate the motion of the vehicle, which can complement the IMU data for more accurate mounting angle estimation. By integrating visual information with IMU data, the model can improve its performance in scenarios with complex motion patterns or when GPS signals are unavailable.

How could the proposed deep learning model be integrated with classical filtering techniques, such as Kalman filters, to further enhance the performance of the overall system

Integrating the proposed deep learning model with classical filtering techniques, such as Kalman filters, can further enhance the performance of the overall system by combining the strengths of both approaches. The deep learning model can provide high-level estimates of the mounting angle based on the sensor data, while the Kalman filter can refine these estimates by incorporating the model's predictions and the sensor measurements in a recursive estimation process. The deep learning model can serve as the prediction step in the Kalman filter, providing initial estimates of the mounting angle. The Kalman filter can then update these estimates based on the incoming sensor data, taking into account the uncertainty in the measurements and the model's predictions. By fusing the outputs of the deep learning model with the Kalman filter, the system can benefit from the model's ability to learn complex patterns from data and the filter's ability to incorporate dynamic system models and sensor noise characteristics for more accurate and robust mounting angle estimation.