Context-Aware Adaptive Wheelchair: AI-Driven Terrain Detection for Mobility Assistance

Abstract

The lack of economical assistive devices limits people with lower limb disabilities in terms of mobility, autonomy, and overall life satisfaction. Powered wheelchairs do offer advanced wheelchair functions for users but are very expensive and rely on sophisticated technologies to supplement their usage. Due to this, many use manual wheelchairs, which require a significant amount of effort to self-propel, especially on uneven surfaces. This paper presents an augmentation of a standard manual wheelchair with context-aware assistance that helps in difficult terrains. The system uses EfficientNet, a lightweight and accurate deep learning model for classification which improves the accuracy of classification by up to 2.5% compared to single scaling methods. To try out our method, we test EfficientNet against ResNet50 and MobileNetV3 to compare accuracy, inference speed, and computational efficiency. Depending on the classified terrain, which includes ramps, side slopes, or rough pavements, the system suggests whether powered assistance to help with the rotation of the wheelchair wheels should be provided or not, thus reducing physical exertion while allowing the user to use their arms. This method meets the user’s needs by providing an optimum combination of ease of use, efficiency, and reasonable cost.

DOI: https://doi.org/10.24321/2454.8650.202601