Illumination-Robust Head Pose Estimation for Enhanced Smart Wheelchair Control

Smart wheelchairs based on head pose estimation offer an alternative solution for individuals with limited arm and leg function to aid mobility. However, their performance is significantly affected by illumination conditions, particularly in low-light environments. Most previous studies lack real-world implementation and have only been evaluated in laboratory settings. To address these limitations, this research investigates the implementation of image preprocessing techniques to enhance the robustness of head-controlled wheelchair systems under different lighting conditions. This study presents a comparative analysis of three image preprocessing techniques: Histogram Equalization, AutoContrast, and DeepISP, evaluated on an actual smart wheelchair prototype. Experimental results show that DeepISP consistently achieves the highest accuracy, reaching 0.942 in normal-light and 0.933 in low-light environments, resulting in more stable and reliable head movement detection. Although DeepISP requires slightly higher processing time (34.82 ms), the latency remains within real-time constraints, representing an acceptable trade-off between accuracy and responsiveness. These findings highlight the importance of illumination enhancement in improving usability, safety, and reliability of vision-based assistive wheelchair systems across diverse lighting conditions.