Trajectory reconstruction of crashes and near-crashes from 100-Car NDS time-series data

This repository reconstructs bird's eye view trajectories of vehicles involved in crashes and near-crashes from 100-Car Naturalistic Driving Study (NDS) radar data.

100Car Naturalistic Driving Study (NDS)

The 100-Car NDS was an instrumented-vehicle study conducted in the Northern Virginia / Washington, D.C. area over a 2 year period in early 2000s [^1]. The primary purpose of the study was to collect large scale naturalistic driving data. To this end, the instrumentation was designed to be unobtrusive, study participants were given no special instructions, and experimenters were not present. Approximately 100 vehicles were instrumented with a suite of sensors including forward and rearward radar, lateral and longitudinal accelerometers, gyro, GPS, access to the vehicle CAN, and five channels of compressed digital video. [^1]: Dingus, T.A., Klauer, S.G., Neale, V.L., Petersen, A., Lee, S.E., Sudweeks, J., Perez, M.A., Hankey, J., Ramsey, D., Gupta, S. and Bucher, C., 2006. The 100-car naturalistic driving study, Phase II-Results of the 100-car field experiment DOT-HS-810-593. United States. Department of Transportation. National Highway Traffic Safety Administration.

From the data collection, an event database has been compiled for 68 crashes and 760 near crashes, as defined in the table below [^2]. Note that 75% of the single vehicle crashes were low-g force physical contact or tire strikes; in other words, most of the crashes involved very minor physical contact. [^2]: Neale, V.L., Dingus, T.A., Klauer, S.G., Sudweeks, J. and Goodman, M., 2005. An overview of the 100-car naturalistic study and findings. National Highway Traffic Safety Administration, 05-0400.

Including time-series sensor data, event context narratives, and manually remarked descriptions of these events, this database is now made public [^3] under a license of CC0 1.0. The time-series profile for each event contains radar and accelerometer data spanning 30s before the event and 10s after the event. This allows for trajectory reconstruction for the vehicles involved in the events. [^3]: Custer, K., 2018. 100-Car data. VITTI. https://doi.org/10.15787/VTT1/CEU6RB

Reconstructed examples of crashes

Not all of the events can be reconstructed due to the missing values, inaccuracy of sensing, and the lack of a ground truth. Subsequently, matching the target vehicle among the detected vehicles in each event is neither trivial. In this repository, 11 crashes and 169 near-crashes are matched based on the restriction that there is not sufficient space (4.5 m) for an undetected vehicle. The following example visualises one of the reconstructed and matched crashes. For the rest of the examples, please refer to the folder ./visual_examples.

To repeat/adjust the processing

Python libarary requirements

pandas, pytables, tqdm, numpy, matplotlib

Wrokflow

Step 1. Download the raw data from [^3] in the folder RawData. This include: 100CarVehicleInformation_v1_0.txt, 100CarEventVideoReducedData_v1_5.txt, HundredCar_Crash_Public_Compiled.txt, HundredCar_NearCrash_Public_Compiled.txt, Researcher Dictionary for Vehicle Data v1_0.pdf, Researcher Dictionary for Video Reduction Data v1.3.pdf, and DataDictionary_TimeSeries_v1_2.pdf. (Given that the license of raw data is now CC0 1.0, which means no limits, this repo has included needed data for your convenience.)

Step 2. Convert 100CarVehicleInformation_v1_0.txt into 100CarVehicleInformation.csv using microsoft excel or other data sheet tools, and rename the column names based on corresponding data dictionary; similarly, convert the 100CarVehicleInformation_v1_0.txt into 100CarEventVideoReducedData.csv, rename and remain the columns of webfileid, vehicle webid, event start, event end, event severity, target type, event nature, then remove "Conflict with " in the descriptions and rename the column name event nature by target. (This has also been done in this repo.)

Step 3. Run preprocessing_100Car.py

Step 4. Run processing_100Car.py

Step 5. Run event_matching.py, which can be adjusted for your own matching

Step 6. Use visualiser.ipynb to observe the reconstructed events

Copyright

Copyright (c) 2024 Yiru Jiao. All rights reserved.

This work is licensed under the terms of the MIT license. For a copy, see https://opensource.org/licenses/MIT.