Deep Learning Based Smart Wheelchair Navigation Optimization for Multi-Lighting Conditions
Amila Fadhila Rahmaniati, Fitri Utaminingrum · 2024
Smart wheelchairs have begun to be developed to help people with special disabilities in helping their independent mobility, including the use of head movement for the navigation. One of the major challenges in smart wheelchair navigation is low-light conditions, which poor image quality can interfere with automated navigation systems. The purpose of this study is to develop a smart wheelchair navigation system that is able to operate well in low-light conditions, thereby supporting the improvement of the performance in multi-lighting conditions. This study proposes the integration of the low-light enhancement method, RetinexNet, for the preprocessing stage applied to the EfficientNetv2 model to improve the quality of head movement image data in low-light conditions. The proposed approach successfully improves the performance of the classification process, which initially experienced overfitting on low-light images. In low-light conditions, the model evaluation metrics increase significantly with all metric values reaching 1.0 in all model versions (BO, B1, B2, B3), indicating overfitting. The highest performance produced on enhanced images was achieved by the EfficientNetv2 B2 variation, with precision, recall, and f1-score values of 0.735, and accuracy of 0.873. The use of the RetinexNet method for low-light enhancement has been proven to improve image quality, which also affects the classification performance of the EfficientNet model in multi-lighting conditions.