By train or by car? Detecting the user's motion type through smartphone sensors data

Bedogni, Luca; Di Felice, Marco; Bononi, Luciano · 2012 · Crossref

DOI: 10.1109/wd.2012.6402818

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of automatically detecting a user’s motion type (walking, driving by car, or traveling by train) using smartphone sensor data. While determining location via GPS is straightforward, identifying the specific mode of transport is more complex but valuable for context-aware applications, such as automatically adjusting device settings or providing smart home services based on estimated arrival times. The authors aim to demonstrate that smartphone sensors can reliably distinguish between these motion types and to optimize the classification process for accuracy and energy efficiency. The study combines experimental data collection with analytical modeling. The authors developed an Android application to sample accelerometer and gyroscope data at a fixed rate of 10 Hz. They collected approximately 72,000 samples per motion type from various users in heterogeneous environments, allowing for natural device orientation changes. To handle orientation variability, they computed an orientation-independent "magnitude" metric for sensor vectors. Features extracted from 10-second time sequences included the minimum, maximum, average, and standard deviation of the sensor magnitudes. The researchers compared three supervised machine learning algorithms: Random Forest, Support Vector Machines (SVM), and Naive Bayes. They also analyzed the impact of sampling rates, sequence lengths, and sensor combinations on performance. The results indicate that combining data from both the accelerometer and gyroscope significantly improves classification accuracy compared to using either sensor alone. Among the algorithms tested, Random Forest achieved the highest overall accuracy at 97.71%, outperforming SVM (82.50%) and Naive Bayes (84.11%). Specifically, Random Forest achieved 96.70% accuracy for walking, 94.10% for driving, and 98.80% for train travel. The study found that sequence length had little impact on accuracy, allowing for shorter sequences to increase responsiveness. Furthermore, lower sampling rates were sufficient for high accuracy, offering a trade-off that reduces energy consumption. To mitigate random classification errors, the authors implemented a history-based filtering mechanism that uses the majority vote from recent predictions to stabilize the output. The authors integrated these findings into an Android application called "What Am I Doing" (WAID). This application runs in the background, detects the user's motion type, and allows users to associate specific device profiles (e.g., silencing ringtones) with each motion type. The detected context is exported via an Android Content Provider, enabling other applications to access this information. The paper concludes that multi-sensor fusion and Random Forest classification provide a robust method for motion type recognition, paving the way for more sophisticated context-aware mobile services and smart city applications.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction