Mode of Transportation Classification Using Android Sensor Data
Anwar, Muhammad Sulaiman (2026)
2026
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-05-29
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202605296538
https://urn.fi/URN:NBN:fi:tuni-202605296538
Tiivistelmä
Automatic transport mode detection, identifying whether a person is travelling by bicycle, boat, bus, car, or tram, plays a key role in sustainable urban planning, public transport optimization, and carbon emission monitoring. However, most existing systems rely on GNSS-derived position and velocity, which require at least four simultaneously visible satellites and fail in challenging environments such as tunnels, underground stations, and dense urban canyons.
This thesis proposes and evaluates an approach that bypasses explicit location estimation and instead utilizes raw Global Navigation Satellite System (GNSS) signal observables along with Inertial Measurement Unit (IMU) data collected from Android smartphones. The GNSS receiver provides three primary observables per visible satellite at each epoch: pseudorange (time-of-flight distance estimate), carrier-to-noise density ratio (C/N₀), and Doppler frequency shift (relative satellite-receiver velocity). To reduce common error sources such as satellite clock drift and atmospheric delays, Single-Difference (SD) and Double-Difference (DD) transformations are applied across time epochs while preserving motion-related variations. The IMU contributes tri-axial accelerometer, gyroscope, and magnetometer readings. A 10-second rolling window is used to extract statistical features, including mean, standard deviation, median, median absolute deviation, minimum, and maximum. This results in 24 GNSS features, 54 IMU features, and 78 fused features per sample. Three configurations are evaluated: GNSS-only, IMU-only, and GNSS+IMU fusion. Histogram Gradient Boosting (HGB) is selected as the classifier due to its ability to handle high-dimensional tabular data, its scale invariance, and its iterative error-correction mechanism.
Experiments are conducted over five transport modes: Bike, Boat, Bus, Car, and Tram across eight classification scenarios ranging from three-class to five-class problems. A strict file-wise split is used to prevent data overfitting by ensuring that complete journey recordings are assigned exclusively to either training or testing sets. A critical finding is the impact of evaluation methodology: random sample-wise splitting artificially inflates accuracy to approximately 99% due to data overfitting, whereas the correct file-wise evaluation yields realistic performance between 38% and 88%. This highlights the importance of robust evaluation design in transport mode classification research.
This thesis proposes and evaluates an approach that bypasses explicit location estimation and instead utilizes raw Global Navigation Satellite System (GNSS) signal observables along with Inertial Measurement Unit (IMU) data collected from Android smartphones. The GNSS receiver provides three primary observables per visible satellite at each epoch: pseudorange (time-of-flight distance estimate), carrier-to-noise density ratio (C/N₀), and Doppler frequency shift (relative satellite-receiver velocity). To reduce common error sources such as satellite clock drift and atmospheric delays, Single-Difference (SD) and Double-Difference (DD) transformations are applied across time epochs while preserving motion-related variations. The IMU contributes tri-axial accelerometer, gyroscope, and magnetometer readings. A 10-second rolling window is used to extract statistical features, including mean, standard deviation, median, median absolute deviation, minimum, and maximum. This results in 24 GNSS features, 54 IMU features, and 78 fused features per sample. Three configurations are evaluated: GNSS-only, IMU-only, and GNSS+IMU fusion. Histogram Gradient Boosting (HGB) is selected as the classifier due to its ability to handle high-dimensional tabular data, its scale invariance, and its iterative error-correction mechanism.
Experiments are conducted over five transport modes: Bike, Boat, Bus, Car, and Tram across eight classification scenarios ranging from three-class to five-class problems. A strict file-wise split is used to prevent data overfitting by ensuring that complete journey recordings are assigned exclusively to either training or testing sets. A critical finding is the impact of evaluation methodology: random sample-wise splitting artificially inflates accuracy to approximately 99% due to data overfitting, whereas the correct file-wise evaluation yields realistic performance between 38% and 88%. This highlights the importance of robust evaluation design in transport mode classification research.