Project leader : LYNRED
Associated organizations : CHIPS JU, BRIGHTER project, Neovision
Research focus : Imagerie thermique et IA
Objective: remove barriers for AI researchers by giving them access to a high-quality, royalty-free dataset
Related publication : Dataset web page
Description :
Artificial intelligence is playing a crucial role in society's digital transformation, particularly in the field of computer vision. However, to be truly effective, these systems require training data, essential for building robust AI models. This is why LYNRED is involved in research, making thermal imaging datasets available to push back the limits of what the technology can achieve. In particular, the focus of this dataset is on applications for safer mobility. In the automotive sector, future emergency brake assist systems are likely to be based on infra-red imaging, which remains effective even in poor visibility conditions.
This dataset is divided into three parts covering complementary fields:
- Multimodal detection: for training AI algorithms, with up to nine classes under different conditions
- Stereovision: to combine multimodal, stereo thermal IR and stereo RGB visible light, as well as tracking (video sequences), with perfectly synchronized images
- Range estimation: for estimating pedestrian detection range under various automatic pedestrian emergency braking (AEBB) conditions, which go beyond current regulations
Multimodal detection
LYNRED's multimodal sensing dataset is specifically designed for the development of advanced driver assistance systems (ADAS) and autonomous vehicles, while providing a collection of thermal and visible RGB data. It comprises 8,000 synchronized infrared and visible RGB images captured in a wide range of conditions, including all seasons, day and night. Researchers and engineers are invited to use this dataset to develop and test their algorithms under real driving conditions.
Stereovision
LYNRED's Stereovision dataset includes all the elements needed to develop algorithms for image registration, visible-thermal fusion and depth estimation. It contains 43,200 images from six video sequences, using two thermal and two visible cameras synchronized by an external trigger in different urban, rural, day and night scenarios.
Range estimation
LYNRED's range estimation dataset provides a large number of sequences of pedestrians crossing the road at multiple distances, captured from a fixed camera inspired by scenarios from the New Car Assessment Program (NCAP) and the National Highway Traffic Safety Administration (NHTSA), two renowned automotive safety performance evaluation programs. It can be used to evaluate the detection range of thermal PAEB systems up to 250 meters, as well as other applications with variable range in relation to pedestrian detection such as autonomous vehicles and ADAS, which exceed the range of NCAP and NHTSA test protocols.
