Robert Spang studied computer science (B.Sc.) at the Technical University of Berlin and worked as a software developer for several years. He then moved to Scotland to study psychological science (M.Sc.) at the University of Glasgow (2018). Later that year, he joined the Quality and Usability Lab at TU Berlin as a PhD student to work on research projects that help and support people. His research interests range from user experience, cognitive psychology, and machine learning to biofeedback-based behavior prediction. As part of a project with Charité Berlin, he is working on the analysis of mobility data to study mobility in old age.
Motivation & Contribution
Mobility researchers using GPS first obtain raw coordinates and timestamps from GPS instead of the variables they're interested in. Conversion is needed to acquire, for example, the time spent out of home, the number of revisited places, or the total time spent on the go. All of these rely on the ability to precisely identify stops and trips and are therefore fundamental when it comes to mobility research.
The commonly adopted strategy involves a combination of a distance and a time threshold to identify significant places (Ash- brook and Starner, 2002, Ye et al., 2009). Here, GPS records are grouped together, if they lie within such a pre-defined radius and time. When we planned the technical basis for a mobility intervention study, we tested several existing systems based on this approach. We observed, on the one hand, significant segmentation of the identified stops, due to the relatively large amount of signal noise. On the other hand, we could only identify stops having a duration greater than a pre-defined time threshold, usually five minutes. Hence, the temporal resolution of this analysis was sub-optimal. Reduced this threshold, lead to an increased number of falsely identified stops (false positives) and segmentation. To solve this, we developed a modern stop and trip identification algorithm.
For a human annotator, this task is fairly easy: when dwelling on a spot, the GPS records scatter around the true position because of its imperfect signal. Records obtained from a trajectory through an environment are clearly distinguishable - although the imperfect signal diverges from the true position similarly. This observation inspired us to create a new algorithm around the idea of investigating the signal patterns, and therefore the geometric properties of the signal noise.
We describe the algorithm's mechanics in detail and discuss its design decisions. Further, we provide benchmark results against established and frequently cited libraries.
Fundamentally, the algorithm is based on a multitude of different, geometric analyses. Each analysis method is applied to a rolling window of subsequent GPS samples. For example, one metric evaluates the ratio between total path length and the bounding box of the set. Another is concerned with the mean angles between the point vectors. Subsequently, all metrics are combined to form a majority-based classification decision for each individual GPS sample. This way, the different methods can compensate for a wrong decision of a minority of the metrics.
If available, the acceleration of the device is also taken into account to exclude unambiguous periods of non-movement. Therefore, we created a simple metric that transforms a three-dimensional vector of x, y, and z acceleration into a motion score that expresses the amount of physical movement of the recording device.
The labels of individual GPS samples are then used to aggregate stop intervals. In the last step, the resulting stop intervals are filtered. Therefore, each interval is compared against the neighboring ones to decide if a) it should be kept as it is, b) if it should be merged with a close stop-interval to reduce segmentation, or c) if it should be discarded.
To test the accuracy of our analysis approach, we benchmarked the system against the built-in methods for stop and trip detection of Moving Pandas (Graser, 2019) and Scikit Mobility (Pap- palardo et al., 2019). These represent a large share of the most commonly used tools for mobility research.
To test the classification performance, we created a large dataset containing trajectories from over 126 days of everyday life and captured 692 stops.
This reference acts as ground truth for the comparison of different frameworks. We investigate sample-by-sample classification metrics (accuracy, precision, recall/sensitivity, specificity, and F1) and stop/trip interval specific metrics (stop-counts, several metrics to quantify the number of detected stops against the reference, such as % matched reference stops, absolute duration error, missed stop duration, absolute start deviation, absolute end deviation, and position deviation). To ensure a fair comparison of the algorithmic approaches, we did not take the acceleration data into account, as the reference systems do not support filtering stop and trip intervals using this kind of data.
Results & Discussion
Our Stop & Go Classifier outperforms other systems in most metrics: it identifies more stops correctly, the stops it misses are shorter in duration, and the start and end times of the identified stops are almost twice as precise as the closest competitor.
The core ideas of the system are a) it uses unfiltered, raw GPS data, b) it analyzes these regarding their geometric properties, and c) it uses multiple scoring mechanisms to create one solid classification.
The Stop & Go Classifier is free software under a BSD 3-Clause license. The repository includes a reference implementation of the algorithm and small usage examples: https://github.com/RGreinacher/Stop-Go-Classifier
Motivation & Contribution
Part of the development of an analysis pipeline for mobility studies using GPS data is benchmarking its performance on both the raw data accuracy and the analysis pipeline itself. When we started to develop our algorithm for stop and trip classification, it became clear that we needed a precisely annotated dataset containing accurate stop and trip labels as a ground truth. Apart from validating our development, we wanted to have a reference point for comparing our analysis methods with existing libraries.
For the study, we planned to equip participants with a smartphone to collect movement data in form of GPS and acceleration data for several days in a row. To prolong battery time, we chose a lower sample frequency. Our special focus was to create ground truth for stop and trip detection algorithms, hence the annotation focused on this.
Through this manuscript, we contribute a comprehensive dataset providing accurate start and end timestamps for stops over 126 days. The STAGA dataset is an unprocessed table of GPS coordinates, annotated with a timestamp, altitude, GPS accuracy, and class label ("stop" or "trip"). Each sample labeled as a "stop" further contains the GPS coordinates of the location it's attributed to. The acceleration data is provided as a separate file, but covers the same time frame and contains a triple (x, y, z) of acceleration sensor readings for each given timestamp, sampled at 1 Hz. The STAGA~dataset is provided publicly and free to use. We further provide the iOS app used to create the diary data for simple stop/trip annotation while on the go. All this is made available under CC BY 4.0.
To create the dataset, we first tried a traditional diary approach: four researchers were taking notes, writing down addresses and times whenever they stopped. While this provided some first samples, it was a tedious and error-prone process, since taking notes is impractical in everyday life. Furthermore, it required looking up the coordinates belonging to each noted address, which works for clearly defined, urban spaces but can become problematic otherwise, e.g. in a park or a rural, outdoor environment as addresses aren't precise enough here. Because of that, we developed a simple iOS app that helped us annotate our movements. The app contains a map to validate the identified position, one button to start or end a stop, and a list overview of previously recorded stops. It captures the GPS position whenever a new stop is started and stores the current time as the start timestamp. When the button is pressed again, the stop is completed and the current time is stored as the end timestamp. Trips are derived from the intervals between two stops. Even more, the app allows exporting the captured annotations as a CSV file which can be directly used for benchmarking purposes. This way, we were able to create a GPS dataset containing precise stop/trip annotations, together with a reference position of the actual stop location. The diary was recorded using an Apple iPhone XR.
The device we used for the recordings was a ZTE Blade A5 (2019). It was configured to record GPS samples at a minimum accuracy of 25m, so if the device was unable to obtain a position reading within this radius, the data point was omitted. We sampled data with a frequency of 0.1 Hz and used both network and GPS as sources for determining the position (the smartphone supports A-GPS and GLONASS). It runs Android 9 and is equipped with a 2.600 mAh battery; during the recording of the dataset, the battery was always charged before the phone shut down.
While the dataset contains mostly everyday life, it also holds small periods of vacation, travel, and hiking. Most trips were carried out by bike. However, the dataset contains long periods of walking, car traffic, and train rides as well. While the data was recorded in two different European countries (mostly urban environments), everything was rotated and projected into the North Atlantic for privacy protection. In the same vein, all timestamps have been shifted to start on January first in the year 2000. However, none of these changes should affect the performance of stop and trip detection algorithms, as the relative temporal and spatial accumulation of GPS records are not changed.
The dataset contains 122,808~GPS and 7,813,740~accelerometer records. The recording time spans over 126.65~days.
The diary contains 692~stops and 691~trips.
The average (mean) duration of a stop is $240.8min$; the average trip duration is $22.7min$.
On average, a stop contains $114.0$ GPS samples; a trip contains $63.5$ GPS samples (mean).
Discussion & Use-Cases
This dataset enables researchers to validate the performance of their algorithms that are used to predict stops and trips from GPS data. It provides a ground truth through careful annotations over a long period. In particular, the development of algorithms for stop and trip classification should profit from this dataset as it enables accuracy tests in the temporal and spatial domain. Due to free access, researchers can use it in various projects, enabling them to make data-driven decisions in the development of mobility research frameworks.
The described dataset, containing GPS & acceleration records and stop/trip annotations, are publicly available at the Open Science Framework under a CC-By Attribution 4.0 International license: https://osf.io/34sft/
The annotation companion app we used to annotate the dataset is free software under a BSD 3-Clause license: https://github.com/RGreinacher/GPS-Diary