On the requirements on spatial accuracy and sampling rate for transport mode detection in view of a shift to passive signalling data. (May 2020)

• Record Type: Journal Article
• Authors: Burkhard, O.; Becker, H.; Weibel, R.; Axhausen, K.W.
• Is Part Of: Transportation research. Volume 114(2020)
• Page Start: 99
• Page End: 117
• Publication Date: 2020-05
• Subjects: Transportation mode detection -- Data quality -- Passive tracking; Transportation -- Periodicals; Transportation -- Technological innovations -- Periodicals; 388.011
• DOI: 10.1016/j.trc.2020.01.021
• Languages: English
• ISSNs: 0968-090X

Abstract

Highlights: Segmentation based classification is robust enough for today's passive tracking. Improper training and testing split can significantly distort reported accuracies. Point-based (online) classification underperforms segment based classification. Recurrent Neural Networks yield better label accuracies but worse label sequences. Abstract: GPS based campaigns have been hailed as an alternative to transportation surveys that promise relatively high accuracy at a relatively low burden on the participants and fewer forgotten trips. However they still necessitate the recruitment of participants and are thus potentially biased and certainly not encompassing significant parts of the population. Given the high penetration of mobile phones, passive tracking by telephone providers would alleviate those two shortcomings at the cost of reduced sampling frequency and positional accuracy. The trade-off in quality has not yet been quantified and therefore recommendations on sensible thresholds are not yet available. In this study therefore, instead of presenting yet another method for mode of transport classification, we therefore compare the performance of existing mode detection schemes under deteriorating sampling rates and positional accuracies. As a possibility to compensate for the deteriorating signal we also calculate features from users' positional histories that could be beneficial if their behaviour is repetitive. The evaluation is not only based on pointwise accuracy, but includes quality measures that pertain to trips as a whole. We find that the necessary accuracy and sampling rate for applications will depend on whether the information of whole trajectories can be used, or whether only the current information is available. The former being relevant to ex-post analyses while the latter situation appears more frequently in near-time analyses. For segmentwise classification, there is no major impact on the quality of the classification by the tested levels of spatial accuracies as long as the sampling intervals can be kept at or below a minute, whereas for point based classification the sampling interval should be between 30 s and a minute and increasing spatial accuracy always improves the classification.