This German Traffic Transfer Learning (GTTL)-dataset contains actually more than 16,000 images with more than xx classes. The class system is based on the German road traffic regulation (StVo), traffic and traffic participants. It covers traffic signs, lights, vehicles, pedestrians. The dataset contains city, villages, country roads, federal roads and highway scenes in different weather conditions, daytimes and time of year. It includes a variation of corner cases like extreme light, snow and rain to train and validate Deep Learning algorithms.
This data set will be continuously expanded and maintained by me in the future. The aim is to collect and label a huge amount of images (depending on my private time). Additionally it would be nice to get support, because the prices for gasoline have risen sharply and as a student I pay the driving distances to record the data myself.
Data:
Resolution: 1920×1080 pixels
Camera: RPi IR-Cut Camera (IR-Filter deactivated)
Sensor: OmniVision OV5647
Lens: M12 mount lens (3.6 mm focal length)
Format: PNG (Quality Level 95%, Compression Level 0)
Size: 93.7 GB
Classes: 284
Images: 16.175
Minimum distance between individual images: individually self-triggered
Hardware:
SBC: Raspberry Pi 4B
OS: Raspberry Pi OS (64-bit)
CPU: 1.9 GHz (overclocked)
GPU: 650 MHz (overclocked)
RAM: 8GB
EEPROM: up to date (important for optimized speed)
+ specific settings (see RPi 4 Tuning guide*)
1x Raspberry Pi 4 Model B 8GB RAM
1x Fan SHIM by Pimoroni
1x RPi IR-CUT Camera
1x 3D mount with GoPro bracket
1x 10.1inch Capacitive Touch Display for Raspberry Pi, 1280×800, IPS, DSI Interface
1x IP2S-S3L Car Tablet Holder
1x CSI Flex ribbon cable
1x Odoga 300W Digital Power Converter
The following image shows the Raspberry Pi 4 with a heat sink being machined to be compatible with the fan.
RPi 4 Tuning guide*. Of course the values are not limited, but for me the focus was on a combination of performance and safety. If necessary, I will adjust the values for the CPU and GPU. All the other components can easily be found on the web.
Here is my setup I used to record the data during driving on different scenes. The used components can be found in chapter Setup.
The Computer Vision Annotation Tool (CVAT) is an interactive image and video annotation tool for computer vision. It supports faster labeling than conventional tools with functions like automated annotation, tracking mode or pretrained AI models. So please save yourself time and use a advanced tool instead of the slow tools.
The Plate Recognizer is an Automatic License Plate Recognition to detect and blur license plates of private vehicles. The software also is able to detect human faces and to blur them. Since German data privacy prefer to make these two things unrecognizable, I used this tool to quickly and easily get a dataset that can be published without problems. With a few hundred exceptions, blurring faces and license plates was not a problem.
Note: the images will be converted to .jpg format.
Part 1: countryroad and village
https://mega.nz/folder/qmIDkKiY#HDYF0MwsnrmXNzCrG7a3hQ
https://mega.nz/file/emgQnTDB#0LHeJnB2mXnvBINEj82qYw_XE7mO5YC7oE-FaIfh1do
Part 2: highway in progess
Part 3: town in progress
I would be glad, if someone could take the existing data and label the dataset with license plates and faces, so that the community have a simple automated tool (python script with trained model), which detects and blur very fast these two objects. In relation to the time to blur unrecognized object manually, which are not recognized by the Plate Recognizer a strong model could help to overcome this.
For automated camera triggering to create the dataset, GPS triggering would be a good method. The Raspberry Pi with a GPS module (e.g. Neo-6M GPS with pynmea2 or geopy library could be used for this. The calculation of the distance travelled can be achieved manually pynmea2 or via ready-made functions geopy. The question arises after how many metres the camera should trigger, are there statistics after how many metres on average in road traffic one traffic sign follows the other or does one refer to already created data sets, such as that of EvoTegra, which used a distance of 4m per triggering. In itself, it might be a good idea to start with a trigger distance of 4 m.