Description
Modern smartphones are widespread and play a critical role in daily life worldwide. These devices contain motions sensors such as inertial measurement units (IMUs) that include accelerometers, gyroscopes, and magnetometers. The IMU sensor data can be used in a variety of applications, one of which is identifying different motorized transport modes. Among these sensors, accelerometer provides the most important features for machine learning models used for transport mode detection. Extracting these features requires estimating and removing the gravity component from acceleration data with a method that is not sensitive to sustained linear acceleration. This is especially relevant for motorized transport that typically contain prolonged acceleration and turning periods. Current gravity estimation methods tackle this challenge by using a combination of accelerometer and gyroscope sensor fusion and opportunistic identification of stable motion periods. These methods have been shown to outperform pure accelerometer or sensor fusion based methods. However, due to the noisiness of the real-world motion environment and inaccuracies of low cost sensors, more research is needed to evaluate the algorithm and improve its robustness.In this thesis, a sensor processing pipeline is developed with the goal to implement and evaluate a state-of-the-art gravity estimation algorithm. The pipeline includes data preprocessing, gravity estimation, extraction of linear acceleration and evaluation of the algorithm based on the distance calculated by integrating the tangential acceleration component. Each step of the pipeline is based on a literature review from the field of IMU based motion sensing and attitude estimation. The gravity estimation algorithm is evaluated with both simulated synthetic data and real-world data collected with smartphones. The real-world data contains several hundreds of hours of labeled smartphone sensor data from multiple transport modes.
The results obtained with the simulated data, show that the algorithm performs as expected in controlled and ideal environment. Moreover, the results reveal some of the challenges with the algorithm, such as the sensitivity to gyroscope bias and false keypoints. The results of individual runs with real-world data show that the algorithm correctly captures linear acceleration and turning periods. However, it additionally indicates that the evaluation metric does not always correlate with correct identification of linear acceleration peaks. The aggregated results with real-world data suggest that the algorithm outperforms sensor fusion methods but has a larger median error compared to accelerometer-based baseline methods. When the results are analyzed in terms of transportation mode, it can be seen that road transport modes are less accurate and more subject to error accumulation during long segments than rail transport modes. Finally, the thesis addresses the challenges and highlights the possible future work for the continuation of this research.
| Period | 2024 |
|---|---|
| Examinee | Erik Oskari Nihtilä |
| Examination held at |