SIMUSAFE cyclist behavior in simulator and in real-world

This dataset includes data collected by the SIMUSAFE H2020 EU project (2017-2021) during its first data acquisition cycle. Voluntary bicycle riders are the subjects in this dataset, and the dataset includes a combination of sensory and psychological characteristics data. Sensory data was recorded in one of two settings: driving around the city in reality (NDT) and driving in a simulator (NST) along routes that were designed to imitate similar in-city driving.

Overall, the dataset consists of recordings from 7 subjects for the following durations:

Total time per user (hours):

User                        Overall recording duration

USER0                      0 days 17:28:06.284997888

USER1                       0 days 04:25:50.084999680

USER2                       0 days 00:11:50.677999616

USER3                       0 days 01:30:38.754000640

USER4                       0 days 02:44:50.625000192

USER5                        0 days 00:24:49.070000128

USER6                        0 days 01:15:46.702999040

Data measurements and computed attributes:

GPS coordinates, acquired by a real GPS receiver in NDT and via a simulated receiver in NST. We extracted velocity from the GPS measurements, computed as the distance between every two subsequent coordinates divided by their corresponding timestamps. As a second derivative, Acceleration was then also derived from the difference of the above-mentioned velocity change between the two subsequent points divided by the time-delta.

Accelerometer data were used to compute the Euclidean norm of the acceleration (a.k.a l2-norm) over the acceleration coordinates vector (i.e., {ax; ay; az}) at each point in time. This feature is sometimes also referred to as the energy-expenditure of the motion.

Additional features:

De/Acceleration {high / low / none}, computed per user per scenario. For each user, acceleration measurements were partitioned by quartiles and were computed per scenario. High-Acceleration was defined as values above the 3rd quartile and low-acceleration as values below. No-acceleration was denoted for the case of acceleration is equal to zero. Respectively, decelerations were computed in an equivalent manner, computed from the partitioning of negative acceleration values.

Data preparation & preprocessing

GPS coordinates were de-duplicated w.r.t subsequent entries.

To avoid issues originating from weak/loss of GPS signal, entries were partitioned into sessions. A session is defined as a sequence of entries with time-deltas no larger than 10 seconds. Velocity & acceleration were derived based on time-deltas within sessions.

Rows with a velocity above or equal to 50km/h were filtered out based on the assumption that a regular bike rider won't reach such speeds.

In addition to the data sources and processing procedures mentioned above, the data has been processed according to the following. Per each subject & scenario, the data was partitioned into windows of 30 seconds using a sliding window with overlap. On each window 5 statistics were computed I.e., entropy, mean, variance, skew, kurtosis on 3 different sources: GPS-based velocity, GPS-based acceleration, and accelerometer-based magnitude. Achieving a total of 15 features.

Data Schema

The data comprises measurement data, data computed after windowing as well as subject psychometric evaluation data. Window data is computed with a sliding window of size 40 (samples) with an overlap of 20 samples. Before computing windows, the measurement data is filtered from entries with missing ‘v_gps’ (velocity computed from GPS coordinates) values.

The measurement dataset is in the attached bicycle_cycle_1_measurement_data.csv file.

Dataset computed with windowing is in the attached bicycle_cycle_1_windowed_data_w_computed_features.csv file.

The psychometric evaluation dataset is in the attached bicycle_cycle_1_subject_psychometric_evaluation.csv file.

Description of all dataset attributes in all three datasets is detailed in the Data description.docx file.