← News

Models & toolsSources: arXiv · 18 Nov 2026

Researchers release WCID benchmark for classifying wheelchair types in automated-vehicle work

A new image dataset labels wheelchairs by configuration, not just presence, and tests how well vision models trained elsewhere transfer to it.

In plain words

Automated vehicles may need to recognize not just that someone uses a wheelchair, but what kind, since chairs differ widely in frame, propulsion, and appearance. The authors introduce WCID, a labeled image collection built for evaluating such recognition, and test several neural networks on it. The models handle clear images but weaken when images are blurred.

What was announced

The paper argues that accessible automated vehicles must account for vulnerable road users, including wheelchair users, and that tailoring the experience requires knowing a chair's specific configuration rather than only that a wheelchair is present. Existing mobility-aid datasets label presence alone, so the authors introduce the Wheelchair Configuration Image Dataset (WCID) to fill that gap. WCID contains 467 face-blurred images across 192 capture groups, intended solely for evaluation. Each image carries a manual-versus-powered label plus annotations for chair subtype, occupancy, viewpoint, setting, lighting, occlusion, propulsion, and control, split into 273 validation and 194 test images.

To measure cross-dataset generalization, the authors train three convolutional neural network backbones and a frozen vision-language probe on an external image dataset, then evaluate them on WCID. They also test robustness with Gaussian blur. According to the abstract, the convolutional models perform well on clean images, but performance drops under strong blur, which the authors describe as leaving a gap before deployment. Code, split assignments, labels, trained weights, and the face-blurred images are available in a public repository.

What it does not show

  • The source is the paper's abstract only, so methods, results, and claims cannot be verified against the full article.
  • The abstract reports no accuracy figures, so the size of the blur-related performance drop is unknown.
  • The dataset is described as for evaluation only, so it does not itself establish how well such models would work in deployed vehicles.

For security and IT teams

Scope
WCID is intended solely for evaluation, with training done on an external image dataset, so it is a benchmark rather than a training corpus.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.