The Stop & Go Classifier

Accurately identify significant places (stops) in GPS trajectories!

Report Bug . Request Feature

Table Of Contents

• About the project
• Installation
• Usage
• Cite this project
• Roadmap
• Contributing
• Author & License
• Acknowledgments

About the Project

The Stop & Go Classifier takes a list of raw GPS samples and identifies significant locations (stops). In other words: It transforms a list of position records into intervals of dwelling and transit.

This is often the first processing step when data-sciencing mobility data. Instead of dealing with raw timestamps and coordinates, one usually wants to see where people went (locations) and how long they stayed (significance). The Stop & Go Classifier is just the right tool for this job.

Key concepts

• Geometric analysis of a GPS trajectory's shape by using the signal's noise and incorporating its properties into the classification decision
• Four independent analyses to form a majority-based decision on how to classify each GPS sample
• Free Python3 software

We provide a complete (open access) paper describing all concepts if you're interested in the nitty-gritty details of how this works. There we also provide a benchmark against well-known Python libraries for stop and trip detection.

🚀 Performance

As the performance values in the paper may become outdated as the code evolves, we provide an up-to-date version of the benchmark. Here, the performance of the Stop & Go Classifier is compared against Scikit Mobility (v.1.2.3; minutes for a stop = 2.167min, spatial radius = 52m) and Moving Pandas (v.0.9rc3; min duration = 164s, max diameter = 114m) using the STAGA-Dataset.

	Scikit Mobility	Moving Pandas	Stop & Go Cl. w/o Motion Score	Stop & Go Cl. incl. Motion Score
% Correct	94.68%	93.34%	96.79%	97.4%
Bal. accuracy	.936	.931	.966	.965
F1 score	.952	.948	.974	.975
Correct stops	94.68%	93.9%	96.79%	97.4%
True stops	74,658	74,047	76,325	76,810
False stops	3,286	3,367	1,571	1,925
Correct trips	92.52%	92.34%	96.42%	95.62%
True trips	40,656	40,575	42,371	42,017
False trips	4,199	4,810	2,532	2,047
-	-	-	-	-
Stop count	1,027	1,407	708	667
Missed stops	34	54	26	26
Fragmented s.	139	234	43	19
Trip count	1027	1407	708	667
Missed trips	32	19	28	33
-	-	-	-	-
Runtime	2.6s	53.17s	49.31s	33.12s

As the Stop & Go Classifier consists of several different analysis methods, we tested the performance of each method independently. We also present the performance change when all algorithms’ results are combined and showcase the influence of the Motion Score and the missing data analysis.

Method	Balanced Accuracy	F1 Score
Width distance ratio	.941	.948
Bearing analysis	.93	.945
Start-end-distance	.948	.955
Path segment intersections	.907	.927
-	-	-
All 4	.962	.968
All 4 + motion score	.963	.971
All 4 + missing data	.966	.973
All 4 + missing + motion	.965	.975

💾 Installation

At this early stage, the classifier is unavailable via the standard package managers (this will come later!). For now, please clone this repository and import the StopGoClassifier.py file.

import sys
sys.path.append('path/to/cloned/repo')
from StopGoClassifier import StopGoClassifier

Dependencies

• scipy>=1.8.0
• numpy>=1.22.3
• pandas>=1.4.1
• (geopandas>=0.9.0 if you need to project raw GPS coordinates first)

There might be earlier compatible versions of the dependencies.

⌨️ Usage

Use the Stop & Go Classifier from StopGoClassifier.py the following way:

# create instance
classifier = StopGoClassifier()

# read input
classifier.read(data.ts, data.x, data.y)

# start pipeline
identified_stops_df = classifier.run()

Note that the classifier expects a planar projection of your coordinates, not the plain GPS longitude/latitude. The example folder provides a demo script to convert one into the other. Other examples cover basic usage including a demo dataset and a simple plot script to display samples and detected stops.

The run() method capsules the following calls:

• process_samples() - classifies each sample as trip or stop
• aggregate() - groups subsequent trips and stops together and forms a table of stops with a start and end time property
• filter_outliers() - decides to either remove, merge, or keep each identified stop

After executing run(), the classifier object offers several interesting variables:

• samples_df - list of all individual GPS samples, including scores from the classification methods and stop/trip labels
• stop_df - list of all stop intervals (the same list that is returned when calling run())
• trip_df - list of all trip intervals, the negative of the stop_df
• trip_samples_df - list of all samples within trip intervals
• debug_stop_merge_df - a list of stop intervals before the merge is applyed. It offers scores of the merge decision methods and is helpful to debug merge-related parameters

The system can be tuned using the following settings:

• MIN_STOP_INTERVAL - time in seconds; stops shorter than this will be ignored
• MIN_DISTANCE_BETWEEN_STOP - distance in meters; minimimum distance between two consecutive stops
• MIN_TIME_BETWEEN_STOPS - time in seconds; remove or merge if less than this threshold
• RELEVANT_TIME_BETWEEN_STOPS - time in seconds; a trip between two stops is relevant if it is longer than this threshold
• MAX_TIME_BETWEEN_STOPS_FOR_MERGE - time in seconds; will not merge stops having more than this time between each other

However, several other parameters, e.g., to disable certain classification methods, are available. These should be described in detail in a wiki. You can provide alternative settings during the classifier's initialization using the optional argument overwrite_settings.

settings = {
	'USE_METHOD_ISA': False,
	'MIN_STOP_INTERVAL': 79,
}
classifier = StopGoClassifier(overwrite_settings=settings)

This repo comes with a few examples and some demo data. Check out the examples folder and run the scripts.

🎓 Cite this Project

This algorithm was introduced at the FOSS4G 2022 conference in Florence, Italy. There, we presented a paper describing the algorithm's architecture and a performance comparison against SciKit Mobility and Moving Pandas's significant locations detection. If you're interested in how the Stop & Go Classifier works, read this paper:

Spang, R. P., Pieper, K., Oesterle, B., Brauer, M., Haeger, C., Mümken, S., Gellert, P., Voigt-Antons, J.-N., 2022. Making Sense of the Noise: Integrating Multiple Analyses for Stop and Trip Classification. Proceedings of FOSS4G, Florence, Italy.

@article{spang2022stopgofoss4g,
  title={Making Sense of the Noise: Integrating Multiple Analyses for Stop and Trip Classification},
  author={Spang, Robert P. and Pieper, Kerstin and Oesterle, Benjamin and Brauer, Max and Haeger, Christine and Mümken, Sandra and Gellert, Paul and Voigt-Antons, Jan-Niklas},
  journal={Proceedings of FOSS4G, Florence, Italy},
  year={2022}
}

(This publication will be available from August 24th, 2022 onwards.)

🗺️ Roadmap

See the open issues for a list of proposed features (and known issues).

🤝 Contributing

Contributions make the open-source community a fantastic place to learn, inspire, and create. Any contributions you make are greatly appreciated.

• If you have suggestions for adding or removing projects, feel free to open an issue to discuss it or directly create a pull request after you edit the README.md file with necessary changes.
• Please make sure you check your spelling and grammar.
• Create individual PR for each suggestion.
• Please also read through the Code Of Conduct before posting your first idea as well.

Creating a Pull Request

1. Fork the project
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a pull request

🙋🏻 Author & 📝 License

👤 Robert Spang, QULab, TU Berlin
🐦 @RGreinacher
✉️ spang➰tu-berlin.de

Copyright © 2022 Robert Spang
This project is BSD 3-Clause licensed.

🙏🏻 Acknowledgments

Thanks for the README generator, Shaan Khan!