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

Abstract

The wide adoption of location-enabled devices, together with the acceptance of services that leverage (personal) data as payment, allows scientists to push through some of the previous barriers imposed by data insufficiency and privacy skepticism. The research problems whose study require hard-to-obtain data (e.g., transportation mode detection, service contextualization, etc.) have now become more accessible to scientists because of the availability of data collecting outlets. One such problem is the detection of a user's transportation mode. Different fields have approached the problem of transportation mode detection with different aims: Location Based Services is a field that focuses on understanding the transportation mode in real-time, Transportation Science is a field that focuses on measuring the daily travel patterns of individuals or groups of individuals, and Human Geography is a field that focuses on enriching a trajectory by adding domain-specific semantics. While different fields providing solutions to the same problem could be viewed as a positive outcome, it is difficult to compare these solutions because the reported performance indicators depend on the type of approach and its aim (e.g., the real-time availability of Location Based Services requires the performance to be computed on each classified location). The contributions of this paper are three fold. First, the paper reviews the critical aspects that are desired by each field of research when providing solutions to the transportation mode detection problem. Second, it proposes three dimensions that separate three branches of science based on their main interest. Finally, it identifies important gaps in research and future directions,i.e., proposing: widely accepted error measures meaningful for all disciplines, methods robust to new datasets and a benchmark dataset for performance validation.

Automatic Trip Detection with the Dutch Mobile Mobility Panel: Towards Reliable Multiple-Week Trip Registration for Large Samples

“…In the past, inaccurate GPS measurements and heavy energy-consuming GPS measurements have been obstacles for large-scale applications (Rieser-Schüssler and Axhausen, 2014). Recently, with increased accuracy and employment of battery-management strategies, a growing number of studies have reported successful pilot applications (e.g., Reddy et al, 2010;Nitsche et al, 2012;Geurs et al, 2015;Cottrill et al, 2013;Zhao et al, 2015;Prelipcean et al, 2014Prelipcean et al, , 2017Safi et al, 2016).…”

Automatic physical activity and in‐vehicle status classification based on GPS and accelerometer data: A hierarchical classification approach using machine learning techniques

“…Using heterogeneous sensor datasets to classify in‐vehicle status and PA with hierarchical processes is novel and has considerable potential for application in a wide range of domains. For instance, Prelipcean, Gidyfalvi, and Susilo () highlighted that interdisciplinary solutions regarding travel‐mode detection should not be limited to one research domain. They also emphasized that the current research trend of validating new algorithms and using datasets that cannot be shared widely hinders interdisciplinary studies.…”

MEILI: A travel diary collection, annotation and automation system

“…Applying AI and ML methods for automatically inferring the aforementioned entities and attributes is an active field of research (see Prelipcean et al, 2014Prelipcean et al, , 2016 for automatic travel mode detection, Bohte & Maat, 2009 for purpose inference and Prelipcean, 2016;J. Wolf, 2006 for destination inference) that does not have any widely accepted methods to perform the inferences or to measure the performance of the methods (Prelipcean et al, , 2017b. In its current form, MEILI uses standard AI and ML algorithms, i.e., a variation of a clustered nearest neighbor classifier for travel mode detection (which obtained an initial precision of 53.5% for 15 travel modes and increased to 75% precision after each user annotated the triplegs of their first days), and a Naive-Bayes classifier for destination and purpose inference (a precision of 54.7% for 13 purposes, and a precision of 41.9% for destination inference 2 ) .…”

“…The learning power of the classifier from annotated data is shown in Figure 7. Whereas the common research focus is infering travel modes (Gong et al, 2017;Prelipcean et al, 2014Prelipcean et al, , 2016Prelipcean et al, , 2017bShafique & Hato, 2017;Zhou et al, 2017) or destinations and purposes (Gong et al, 2016(Gong et al, , 2017Prelipcean, 2016;Su et al, 2014;Usyukov, 2017) after the completion of the case study, the main issue with that approach is the lack of focus on user experience improvements. Embedding active learning as part of the annotation process improves the user experience since it minimizes the amount of data the users have to annotate and gives the users the opportunity to correct the output of the classifiers.…”

“…An in-depth analysis of the performance of the classifiers used by MEILI has been extensively covered by Prelipcean, 2016, where the classifier achieves a top 3 precision of 82.1% for 14 different travel modes. For further discussions on travel mode detection from an interdisciplinary perspective, the reader is directed towards Prelipcean et al, 2017b.…”