| Title |
Adagen: Adaptive Generalized Knowledge Transfer Framework For Sensor-Based Surface Classification For Wheelchair Routing |
| ID_Doc |
6156 |
| Authors |
Misra V.; Menon S.S.; Saha S.; Mathur A.; Yu H.; Raychoudhury V. |
| Year |
2024 |
| Published |
SN Computer Science, 5, 7 |
| DOI |
http://dx.doi.org/10.1007/s42979-024-03181-w |
| Abstract |
The main hurdle to wheelchair mobility is the irregularity and unpredictability of the commuting surfaces. Existing research on accessible routing tackles this problem by generating routes using machine learning algorithms on surface-induced vibrational data. These algorithms are improved using crowd-sourced user data. However, as a wide variety of wheelchairs differ in specifications like their size, wheel type, weight, and structure, it is impractical to collect a balanced dataset to account for every specification of a wheelchair. This paper explores the possibility of transferring the knowledge of collected surface data from observed wheelchairs to that of unknown wheelchairs using a novel approach called AdaGen. The proposed AdaGen framework has a self-learning adaptive activation function, AdaAct AF, which has been shown to outperform the state-of-the-art activation functions in shallow and deep networks. The conducted experiments and the results clearly show that a model trained on a manual wheelchair (MW) dataset can be transferred to a power wheelchair (PW) dataset with a 98.8% accuracy. The proposed model results in better performance with more economical resource usage when compared to a model solely trained on the PW dataset. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2024. |
| Author Keywords |
Accessibility; Adaptive activation; Deep learning; Smart cities; Surface classification; Wheelchair |
