Ab1231/city-sample-freeway-speed

City Sample freeway speed dataset

By Djenadi Mohammed.

Synthetic freeway traffic rendered in Unreal Engine 5.8 on the elevated freeway of Epic's City Sample (the Small City map). Every car has its exact speed on every frame, a segmentation mask and a track ID, so it can train a detector, test a tracker and train a speed model from the same footage.

The video is a 20 second 4K render of the same freeway with the ground truth drawn on. Each car's current speed is worked out from where it is 30 frames before and 30 frames after, next to its maximum and average since it came into view and its true speed. It isn't part of the splits below.

There are two versions. Version 2, in v2/, is the one to use. Version 1 is the first set and is described further down.

Version 2

150 cameras instead of 24, and masks that walls and railings hide properly.

Known bug, to be fixed in an update: version 2 was meant to give regular cars random paint colours, but the colours didn't apply, so its cars still have Epic's default paint like version 1. City Sample's paint takes its colour from a built-in colour table and ignored the one I set. That's fixed in the generator now, and version 2 will be re-rendered with the colours and updated here. The cameras, traffic, masks and speeds stay the same, only the images change.

• 150 clips of 20 seconds at 1920 x 1080 and 30 fps, 90,000 frames in all
• 9,000 labelled images for YOLO segmentation, every 10th frame
• 5,281 cars, from 22 to 150 km/h
• About 9.8 GB

Split	Clips	Images	Masks	Cars	Camera height	Field of view	Pitch
Train	110	6,600	71,316	3,845	4.0 to 21.3 m	25 to 80 degrees	-33 to -3 degrees
Validation	20	1,200	11,267	721	4.0 to 17.4 m	28 to 79 degrees	-24 to -3 degrees
Test	20	1,200	12,840	715	4.5 to 17.9 m	30 to 77 degrees	-22 to -4 degrees

Every clip has its own camera, so a model sees 110 different heights, angles and lenses in training. That's meant to help it work on a new camera without much calibration.

After the update, regular cars will get a random paint colour, weighted like real traffic: white 24%, black 20%, grey and silver 15% each, blue and red 8% each, and a little green, beige, orange and light blue. Taxis, the police car, buses and trucks keep their own paint.

A mask now only covers a car where the camera really sees it. Version 1 masks didn't check the city around the cars, so a car behind the divider wall, a railing or a lamp post still got a mask there.

The splits use the same three stretches of freeway as version 1, and the files are laid out the same way under v2/. Clip names start with C2_, like C2_Train_017.

from ultralytics import YOLO

YOLO("yolo26s-seg.pt").train(data="city-sample-freeway-speed/v2/yolo/data.yaml")

Version 1

The first set, in the top folders. It has fewer cameras, Epic's default paint and the mask problem above, so it's mostly here for comparison.

• 24 clips of 60 seconds at 1920 x 1080 and 30 fps, 43,200 frames in all
• 8,640 labelled images for YOLO segmentation, every 5th frame
• 1,923 cars, from 22 to 150 km/h
• About 6 GB

Split	Clips	Images	Masks	Cars	Speeds
Train	16	5,760	59,175	1,213	22 to 150 km/h
Validation	4	1,440	22,792	382	23 to 146 km/h
Test	4	1,440	8,258	328	38 to 147 km/h

Files

yolo/
  data.yaml                       classes and split folders, ready for Ultralytics
  images/<split>/<clip>_<frame>.jpg
  labels/<split>/<clip>_<frame>.txt
tracks/<split>/<clip>.json        every car on every frame, with its true speed
videos/<split>/<clip>.mp4         the whole clip
showcase/showcase_4k.mp4          the video above
v2/                               version 2, with the same yolo/, tracks/ and videos/ inside

The top folders are version 1.

Clip names look like Train_04, and image names add the frame number, so Train_04_00090.jpg is frame 90 of that clip. Frame n of a tracks file is frame n of the clip's video.

YOLO labels

Each line is one vehicle: the class, then its outline as x, y pairs from 0 to 1.

Class	Name	Masks in version 2	Masks in version 1
0	car	48,089	47,949
1	van	13,431	13,809
2	truck	24,661	20,302
3	bus	9,242	8,165

The outlines only cover what you can see, so a car half behind a truck gets half an outline. Vehicles under 40 visible pixels have no label. A few images have no vehicles in view and an empty label file.

data.yaml has no path line, so Ultralytics finds the split folders next to it:

from ultralytics import YOLO

YOLO("yolo26s-seg.pt").train(data="city-sample-freeway-speed/yolo/data.yaml")

Tracks

Each tracks file has the clip's camera at the top and then one list per frame, with a row for every vehicle you can see.

Field	What it is
id	The car's ID, the same on every frame of the clip
class	0 car, 1 van, 2 truck, 3 bus
kmh	True speed on this frame
box	x1, y1, x2, y2 in pixels around the visible part
pixels	How many pixels of the car you can see
lowest	x, y of the lowest visible pixel, roughly where the tyres meet the road
road_m	Centre of the car in Unreal world coordinates, in metres

kmh and road_m are the answers. A speed model shouldn't take them as input, because real video doesn't have them.

The camera fields are x_m, y_m, z_m, yaw, pitch, fov and height_above_road_m. The field of view is horizontal, the camera has no roll and the principal point is the middle of the image. Unreal world axes are X and Y flat and Z up, and yaw turns from X towards Y.

Speeds come straight from the traffic simulation, which moves each car a set distance every frame, so they're exact.

How it was made

Traffic runs on 3 lanes each way with the Intelligent Driver Model. Every car has its own wish speed and changes it now and then, so cars speed up, slow down and queue behind slower ones. Each clip has its own pace and density, from slow packed traffic to a nearly empty fast road.

The 13 vehicle models are Epic's City Sample cars, vans, trucks and bus. City Sample's own traffic was turned off, so every vehicle in a frame has a label.

The masks come from a second render pass where each car writes its own ID into the stencil buffer. In version 2 a pixel only counts when the car's depth matches the scene's depth there, so anything in front of the car hides it.

Each clip has one fixed camera, on a pole beside the road or over it, facing either way. Version 2's cameras are in its table above, and these are version 1's:

Split	Height above road	Pitch	Field of view
Train	5.9 to 14.4 m	-22 to -6 degrees	31 to 61 degrees
Validation	7.8 to 13.7 m	-15 to -8 degrees	47 to 68 degrees
Test	7.6 to 11.8 m	-18 to -10 degrees	31 to 72 degrees

Splits

The freeway is about 1.9 km long. Each split has its own stretch of it, with gaps between them:

Split	Metres along the freeway	Shape
Train	880 to 1,700	A long straight and a gentle curve
Validation	560 to 740	A short straight between two bends
Test	150 to 520	A big curve

No camera position is shared between splits. A camera can still see far down the road, so parts of another split's freeway sometimes show up small in the background.

Known limits

In version 1, a car behind the divider wall, a railing or a lamp post still has a mask over that part, because the mask pass only checked cars against other cars. Most of its cameras look straight down the road, so this hits a minority of masks, but the outlines and the lowest points of those cars are wrong. Version 2 fixes it.

Everything is from one freeway in one city, at one time of day, with no rain or night.

Version 1 has only 16 training cameras, so a model can partly learn those camera angles instead of what camera height and angle do in general.

Cars keep Epic's default paint in both versions for now, so most cars of one model look the same, like many silver sedans and yellow taxis. A model trained on this can miss dark cars: an early one missed a black SUV in a real photo. Version 2 will be updated with random colours.

The cars are moved along the lanes rather than driven, so their wheels don't spin. There's motion blur on the bodies.

At mask edges JPEG smears the colours, so about 3% of mask pixels were noise and got cleaned up. Outlines can be a pixel or two off at the edges.

All of it is rendered. Treat scores on this data as approximate until they're tested on real traffic video.

Licence and credits

My work on this dataset (the renders, labels, tracks, traffic and camera setups) is under CC BY-NC 4.0. You can use and share it for non-commercial work if you credit me.

For any commercial use, contact me first. That covers training models for a product, selling the data or anything made from it, and using it inside a company's work. Open a discussion in this dataset's Community tab and I'll get back to you.

The scenes and vehicles are Epic Games' City Sample, rendered in Unreal Engine.