Accelerometer-based transportation mode detection on smartphones

Abstract

We present novel accelerometer-based techniques for accurate and fine-grained detection of transportation modes on smartphones. The primary contributions of our work are an improved algorithm for estimating the gravity component of accelerometer measurements, a novel set of accelerometer features that are able to capture key characteristics of vehicular movement patterns, and a hierarchical decomposition of the detection task. We evaluate our approach using over 150 hours of transportation data, which has been collected from 4 different countries and 16 individuals. Results of the evaluation demonstrate that our approach is able to improve transportation mode detection by over 20% compared to current accelerometer-based systems, while at the same time improving generalization and robustness of the detection. The main performance improvements are obtained for motorised transportation modalities, which currently represent the main challenge for smartphone-based transportation mode detection.

Transportation mode detection – an in-depth review of applicability and reliability

“…The body of literature related to LBS contains those studies that focus on detecting the transportation mode of a user as close to real time as possible by relying mostly on data provided by GPS receivers (Stenneth, Wolfson, Yu, & Xu, 2011;Biljecki, Hugo, & van Oosterom, 2013), the accelerometer sensor (Hemminki, Nurmi, & Tarkoma, 2013;Yu, Lin, Yu, Chang, & Wang, 2014) or a combination of the two (Reddy et al, 2010;Manzoni, Maniloff, Kloeckl, & Ratti, 2010;Shah, Wan, Lu, & Nachman, 2014;.…”

“…As there are different types of sensors that can provide data suitable for transportation mode detection, researchers have studied the following sensors: GPS receiver (Stenneth et al, 2011), accelerometer reader (Hemminki et al, 2013;Yu et al, 2014) and the two sensors combined (Reddy et al, 2010;Manzoni et al, 2010;Shah et al, 2014;. The main LBS specific approaches are summarized in Table 2.…”

“…Research that focused explicitly on accelerometer only transportation mode detection report a precision of 70% for six transportation modes using decision tree classifiers (Wang, Chen, & Ma, 2010), 80.1% accuracy and 82.1% recall for seven transportation modes using AdaBoost together with Decision Tree classifiers (Hemminki et al, 2013) and 92.5% accuracy for five transportation modes using a Support Vector Machine classifier (Yu et al, 2014). While the data size available for each research group varies (150 hours from 16 users for Hemminki et al, 2013, 12 hours from 7 users for Wang et al, 2010 and an afternoon from 4 users for Yu et al, 2014), the precision and recall values are computed on accelerometer samples, which are labeled periods of a predefined duration where one accelerometer reading can be found either in one period only (tumbling windows) or in multiple adjacent periods (sliding windows) -(8 seconds for Wang et al, 2010, 1.2 seconds for Hemminki et al, 2013, and 10 seconds for Yu et al, 2014) associated to one activity. The validation of the proposed methods is done by allocating a percentage of the collected data set for testing purposes and training on the remaining set (Wang et al, 2010;Yu et al, 2014), or by using a leave one user out cross-validation (Hemminki et al, 2013).…”

Comparative framework for activity-travel diary collection systems

“…Travel diary collection methods and systems 1) Mode Inference: Inferring the transportation mode of a user is a task that is of interest for two research areas: transportation science [2,7,10,13] and Location Based Services (e.g., GeoLife [21][22][23], and others [14,20]). Scientists collect different data sets -such as GPS-only data [21][22][23], accelerometer-only data [7,20], GPS traces fused with accelerometer data [10,13,14] or GPS traces complemented by GIS information [2,15] -that are annotated by users and afterward use machine learning or rule based systems to train classifiers that automatically determine the transportation modes of future data sets. High classification accuracies have been achieved, which are suitable for travel diary generation: e.g., Prelipcean et al [13] 90.8% (seven classes), Reddy et al [14] 93.6% (five classes), Stenneth et al [15] 93.5% (five classes), and Yu et al [20] 90.6% (five classes).…”

Use of acceleration data for transportation mode prediction

“…In a recent study by Hemminki et al (2013), 16 participants from four countries collected accelerometer data spanning more than 150 h and covering six modes of transportation. A mean recall accuracy of 82.4 % was achieved.…”