Detection of assistive devices used by disabled individuals with YOLO

In order to minimize the disparity between individuals with and without physical disabilities, public institutions and private organizations carry out various initiatives. These efforts include installing tactile paving on sidewalks to assist visually impaired citizens in navigating the path and incorporating gently sloping ramps at the beginning and end of sidewalks to accommodate wheelchair users. With advancing technology, in the near future, robots accompanying disabled individuals or other forms of intelligent electronic warning and guidance systems can serve as navigational aids to overcome the challenges encountered by disabled individuals in their daily lives. Such systems should primarily be able to differentiate disabled individuals from others. This study focuses on the problem of detecting whether a person is disabled or not, based on visual images captured through a camera. Visual object detection, combined with the success of object detection models generated through deep learning, has found application in various fields in recent years. Within the scope of this thesis, works were conducted using deep learning-based object detection methods, namely YOLOv7 and YOLOv8, to detect individuals that have mobility aids. To ensure effective detection for the given problem, we created a Helper dataset consisting of 14965 images. This dataset includes the tools used by individuals with mobility restrictions for movement, as well as commonly encountered objects in daily life to ensure the practical performance of the model. The objects in the dataset comprise 5 classes: crutch, wheelchair, bike, walking stick, and walking frame. We utilized 5-fold cross validation to test the performance of the methods and obtained 97.42% average mAP50 value for YOLOv7-Tiny and %97.30 average mAP50 value for YOLOv8n. These results show that YOLO object detection methods can be used in real applications to detect some basic equipment used by disabled individuals.