Deep Learning for Inertial Sensor Alignment

TL;DR

The paper tackles yaw mounting-angle estimation for a smartphone with an IMU fixed to a car, addressing GNSS-free alignment for navigation tasks. It proposes MountNet, a CNN-based regressor trained with synthetically rotated IMU data to learn the yaw angle from IMU windows, enabling real-time on-device deployment. Key contributions include a large, synthetic-label-augmented dataset, a cosine-based loss to handle angular periodicity, and a real-time smoothing pipeline that yields converged estimates within tens of seconds, achieving MAEs around 4° and robust mid-drive angle-change tracking. The approach demonstrates strong generalization from a low-cost external IMU to a consumer smartphone, performing competitively with GNSS-based solutions while avoiding GNSS dependence and enabling faster, GNSS-free sensor alignment in practical applications.

Abstract

Accurate alignment of a fixed mobile device equipped with inertial sensors inside a moving vehicle is important for navigation, activity recognition, and other applications. 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. In this work, a data-driven approach using deep neural networks (DNNs) is proposed to learn the yaw mounting angle of a smartphone equipped with an inertial measurement unit (IMU) and strapped to a car. The proposed model uses only the accelerometer and gyroscope readings from an IMU as input and, in contrast to existing solutions, does not require global position inputs from global navigation satellite systems (GNSS). To train the model in a supervised manner, IMU data is collected for training and validation with the sensor mounted at a known yaw mounting angle, and a range of ground truth labels is generated by applying a random rotation in a bounded range to the measurements. The trained model is tested on data with real rotations showing similar performance as with synthetic rotations. The trained model is deployed on an Android device and evaluated in real-time to test the accuracy of the estimated yaw mounting angle. The model is shown to find the mounting angle at an accuracy of 8 degrees within 5 seconds, and 4 degrees within 27 seconds. An experiment is conducted to compare the proposed model with an existing off-the-shelf solution.

Figures (14)

• Figure 1: Smartphone device on a mount inside a car. Sensor data collected by the smartphone is measured in the sensor frame of reference (in green).
• Figure 2: Top view schematic of the yaw mounting angle, $\psi$, after adjusting for roll and pitch angles with respect to the gravity vector.
• Figure 3: An example of a single sample in our dataset. We show both $x$ (in dotted black), the $5$ seconds of measurement from each of the IMU channels, and $x_{\pi/2}$ (in blue) the same measurements when artificially rotated by an angle of $\pi/2$ about the $z$ axis as described in Section \ref{['sec:data_generation']}. We can see the effect the rotation has on each of the IMU channels (from top to bottom): acceleration in the x, y, and z axes and angular velocity in the x, y, and z axes.
• Figure 4: Data processing and datasets creation flowchart.
• Figure 5: DNN architecture: three convolutional layers are followed by two dense layers with a connecting mean-pooling layer. The MountNet outputs the mounting angle of the smartphone device inside a car.