Accessible Input on Small Touchscreens: Evaluating Deep Learning Architectures for Stroke and Motion Gestures

Abstract: People with upper-body motor impairments often experience difficulties when interacting with small touchscreens and digital devices. To address this challenge, we present a comparative evaluation of representative deep learning architectures for recognizing touchscreen and mid-air motion input, using an open dataset with gestures collected from people with upper-body motor impairments. Across the evaluated architectures, the 1D-CNN achieved the highest recognition performance, with an accuracy of 94.7% for stroke gestures and 84.0% for motion gestures. Our analysis further revealed that gesture recognition performance, represented by accuracy rates, cannot be fully explained by participants’ health conditions or generic WHODAS health and disability scores, suggesting the need for ability-based design approaches.

Conference: ICUSI ’26 – Proceedings of ICUSI ’26, the 2026 International Conference on User-System Interaction, 140-144.

Link: https://rochi.utcluj.ro/icusi/proceedings/articles/3/26.pdf