A novel image dataset for detecting and classifying mobility aid users

Abstract

When it comes to human identification as a computer vision task, artificial intelligence methods require extensive training to achieve good results. Large-scale image databases used for training and testing are easily available, however, disabled people still have poor representation in human datasets, making their visual identification hard to achieve. In this work, we introduce a new dataset based on wheelchair and cane users to fill the gap of in-the-wild images of disabled pedestrians and enable further research in the area. Additionally, we studied the effect of the dataset using transfer learning on state-of-the-art classification and detection models, training with combinations of the five classes available: wheelchair user, cane user, wheelchair, cane, and able-bodied person. Since this is the first work of its kind, we thoroughly analyzed the classification results across various image sizes and certainty thresholds. Furthermore, detection models trained with the new dataset were compared to those trained with a previously published mobility aid dataset through different evaluation metrics. Our results show high precision and certainty for both classification and detection, demonstrating the benefit the dataset has in the identification of mobility aid users and encouraging the inclusion of disabled people in the development of intelligent systems.