MODELLING PEDESTRIAN MOVEMENT USING SMARTPHONE SENSOR DATA

Abstract

The navigation problem has always been relevant in various fields of human activity, and the advent of cheap electromechanical sensors (IMUs) has led to new applied research in the field of inertial navigation. The main task of inertial navigation is to estimate the position of a moving object at each point in time given its initial speed and position, its acceleration and attitude. Various sensors and methods could be used for the task, including sensor fusion, step counting, visual odometry, machine learning.

The article presents a solution to the task of pedestrian navigation in urban environment. In such settings, satellite signals can suffer from dropouts and multipath effects when GNSS signals reflect off surrounding surfaces, causing the receiver to process both direct and reflected signals. Relying only on IMU sensor data to calculate distance travelled is nontrivial as the result heavily depends on the hardware characteristics and noise. That is why the main goal of the research was to solve the velocity drift problem when integrating raw acceleration in inertial navigation process.

To achieve the goal, we focused on the data preprocessing and its effect on the result. The data from the IMU sensors were sampled at high frequency, 53 Hz, resulting in a constant stream of (noisy) data. We had to filter the noise applying Butterworth low-pass filter to acceleration data and Savitzky-Golay filter to orientation data. Known orientation of the device allowed us to convert measurements from the device frame of reference to the global ENU frame. Smartphone acceleration dataset containing 133000 samples and 2500 target speed measurements was used to train an LSTM model that can regress velocity out of acceleration data. We achieved MAE of 0.087 m/s and R² of 0.83 on the test dataset and applied the full algorithm to reconstruct the recorded rectangular trajectory of 77´15 meters. Simple complementary filter was used for merging predicted and integrated velocities, thus eliminating drift. It was shown that despite the quite accurate speed prediction, the final result depends on the orientation in horizontal plane obtained from the magnetometer.