Researchers have introduced a new dataset to support the development of AI-powered sign language recognition systems that can operate inside vehicles, addressing an environment that has received relatively little attention in digital accessibility research despite its potential importance for shared transport.
The In-Car Sign Language Corpus (ICSL) focuses on Brazilian Sign Language (Libras) and was developed to support future research into sign language recognition within confined spaces such as taxis, rideshare vehicles and carpools. According to the researchers, the long-term goal is to improve accessibility and quality of life for travellers who are Deaf or hard of hearing, who use public and shared transport. While the project is still at the research stage, it provides a foundation for developing and evaluating future digital accessibility technologies rather than introducing a consumer product.
Unlike many existing sign language datasets, ICSL was created specifically for the challenges of signing inside a vehicle. These environments can involve restricted movement, changing lighting conditions, partially obscured hands and non-frontal camera angles, making automated sign language recognition more difficult than in controlled laboratory settings.
To support research in this area, the dataset combines two types of recordings. It includes high-precision laboratory motion capture data, which provides an idealised baseline for sign movements, alongside real-world in-car recordings captured using both standard 2D cameras and 3D Time-of-Flight sensors. In total, the researchers recorded more than 1.5 million synchronised frames featuring Libras users across a range of in-car scenarios. The corpus also includes annotations of both lexical signs and non-lexical elements to support the training and evaluation of machine learning models.
The researchers note that their work is intended as a foundation for future studies rather than a finished accessibility solution. They suggest the corpus could support the development of more robust sign language recognition models, help researchers compare recorded signing with synthesised signing avatars, and improve understanding of how sign language technologies perform in real-world environments.
Although the dataset focuses specifically on Brazilian Sign Language, the research could also inform similar work in other countries. Because sign languages differ between regions, the dataset itself would not be directly applicable to Australia, where Auslan is used. However, the methodology for collecting sign language data in real-world vehicle environments could be adapted to support future Australian research into accessible transport technologies.
For more information, please read The In-Car Sign Language Corpus (ICSL) research paper from arXiv.org.