Lateral conflict resolution data derived from Argoverse-2: Analysing safety and efficiency impacts of autonomous vehicles at intersections

Authors: Guopeng Li, Yiru Jiao, Simeon C. Calvert, and Hans van Lint

• This dataset is derived from the open Argoverse-2 motion forecasting data: [https://www.argoverse.org/av2.html#forecasting-link]
• The crossing conflict dataset, 2 metadata files, and the code to read and visualize the data can be downloaded through the 4.TU data platform: [https://data.4tu.nl/datasets/8d6ee0b0-8ed5-43f3-b1c9-7665cc163e87]. Note that the merging conflicts are not included in this online version. The complete dataset is still under preview of the platform and will be made public once the preview is finished.
• If you want to use the complete lateral conflict data, please contact the authors for immediate share.

Licence

Please use this dataset under the licence of CC BY-NC-SA 4.0 as indicated in the dataset link (4.TU).

Notice

The link above is the initial version. When the paper is published, we will immediately upload the final version with merging conflicts and more annotations.

Quick to start

Reading and visualizing the dataset requires the following packages"

python >=3.8
numpy
scipy>=0.17.0
zarr
h5py
matplotlib
tqdm (optional)

• Put the visual_utils.py in the root dir.
• Unzip the downloaded data.zip file to the root dir.
• The dataset is stored in a compact zarr form so even a 'potato' computer can handle this.

Read and visualize data

• Import the necessary packages:

import zarr
import os
from visual_utils import *

folder = 'av' # or 'hv'

root ='./data_3m/'+folder+'/'

log_ids = [name for name in os.listdir(root) if os.path.isdir(os.path.join(root, name))]
print('Number of scenarios: ', len(log_ids))

• Read the data

'''
slices: len = nb_objects + 1, slices[n] and slices[n+1] gives the start/end indices of the n-th object
maps: lanes as NumPy array
type: len = nb_objects, contains 7 numbers with the following meanings:
        -1: Static background
        0: human-driven vehicles
        1: pedestrians
        2: motorcyclists
        3: cyclists
        4: buses
        10: autonomous vehicles
timestep: timestamps in second, timestep[slices[n]: slices[n+1]] give the timestamps for the n-th object
motion: motion state, with 7 dimensions
    motion[slices[n]: slices[n+1]] gives the motion of the n-th object, the 7 features are the following variables in order:
        [x, y, vx, vy, ax, ay, yaw]
        yaw is to the x-axis, between [-pi, pi]
'''
slices, timestep, motion, type, maps = read_scenario(log_ids[2441], root)

• Visualize the data

# Visualize the 2441-st scenario
# the parameter other_road_users = True/False controls whether to plot the surrounding agents (True by default)
# The parameter direction = True/False controls whether to plot the start and end directions of the vehicles involved in the conflict (True by default)
fig, ax = visualize(log_ids[2441], root, other_road_users=True, direction=True)

About metadata file

The metadata files is a table, shown as the example below, for AV-involved scenarios (metafile_av.csv) and AV-free scenarios (metafile_hv.csv), respectively.

log_id	xi_start	yi_start	xj_start	yj_start	typei	xi_end	yi_end	xj_end	yj_end	typej	direction	PET	ifirst	angle_start	angle_end	start	end
2441	-0.028310	0.999599	-0.278388	-0.960469	AV	-0.001506	0.999999	0.998085	0.061850	HV	R-L	3.7	False	162.213724	86.540315	opposite	cross
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...

These tables provide the following information about each scenario:

column	dtype	information
log_id	str	index of the scenario
[xi_start, yi_start]	float	direction vector of the first* agent recorded in the scenario at the start** time
[xj_start, yj_start]	float	direction vector of the second agent recorded in the scenario at the start time
typei	str	agent type of the first agent recorded in the scenario, being one of {'AV','HV','Pedestrian','Motorcyclist','Cyclist','Bus'}
[xi_end, yi_end]	float	direction vector of the first* agent recorded in the scenario at the end*** time
[xj_end, yj_end]	float	direction vector of the second agent recorded in the scenario at the end time
typej	str	agent type of the second agent recorded in the scenario, being one of {'AV','HV','Pedestrian','Motorcyclist','Cyclist','Bus'}
direction	str	whether the second-passing vehicle moved from the left ('L-R') or the right ('R-L') of the first-passing agent
PET	float	post-encroachment-time
ifirst	bool	whether the first-passing agent is the first agent recorded in the scenario
angle_start	float	angle between the direction vectors of the two agents at the start time
angle_end	float	angle between the direction vectors of the two agents at the end time
start	str	whether the two agents ran parallel (P), crossed (C), or ran opposite (O) to each other before reaching the conflict point
end	str	whether the two agents ran parallel (P), crossed (C), or ran opposite (O) to each other after reaching the conflict point
	Notes	* Note that the first agent does not necessarily pass the conflict point first. ** We consider the start time as 5 seconds before the first-passing agent passed the conflict point, or the start of the record if the time before passing the conflict point is less than 5 seconds. *** Similarly, the end time is 5 seconds after the second-passing vehicle passed the conflict point, or the end of the record if the time after passing the conflict point is less than 5 seconds.

One can use the following code to select data in different conditions:

# AV passing the conflict point first
case_av_first = metafile_av[(metafile_av['ifirst'])] # In cases involving AV, the AV is the first recorded agent

# Pedestrian passing the conflict point second
case_ped_first_av_free = metafile_hv[(~metafile_hv['ifirst'])&(metafile_hv['typej']=='Pedestrian')]