Mobility Assistance for the Visually Impaired in Unstructured Environment

With the rapid upsurge in deep learning technologies and their implementation in various areas of computer vision and image processing, a substantial amount of research work has been focused on obstacle detection and mobility assistance for the visually impaired, the accuracy and performance of which are subject to the availability of reliable datasets. However, these works are mostly based on accessible environments that have a standard structure, which is considerably easier to design and implement. These technologies are prone to failure in congested and unstructured environments, and cannot be utilized by the visually impaired people of regions where pathways and road systems are not consistently constructed keeping in mind the mobility of people with disabilities. The aim of this work is to develop a camera-based mobility assistance system that functions in unstructured environments. In this paper, we propose a novel dataset of obstacles commonly encountered by pedestrians on unstructured roads, that we have collected from various unstructured regions and annotated according to appropriate labels. In the proposed system, we have built a computer vision and deep learning based object detection framework which is trained and tested on the curated dataset. We have also created a separate dataset with the help of visually impaired people, to test our model and determine its robustness in real-world scenarios. The model achieves good overall accuracies on the proposed dataset and shows promising scope for safe and reliable mobility assistance for the visually impaired upon large-scale user-testing.