Prototype of an Indoor Pathfinding Application with Obstacle Detection for the Visually Impaired

THIS study presents an initial prototype for a project aimed at assisting visually impaired individuals using deep learning techniques. The proposed system utilizes the You Only Look Once (YOLOv8) algorithm to detect objects tagged as obstacles. Designed for indoor environments, the system employs a CCTV camera and a computer server running the YOLOv8 model. Additionally, the A-star algorithm is used to determine the optimal path to avoid detected obstacles. Video frames are divided into tiles, each considered a node; nodes with detected objects are marked with a value of 0. The YOLOv8 model currently achieves an initial accuracy rate of 70%, with a mean Average Precision (mAP) at an Intersection over Union (IoU) threshold of 0.5 reaching 0.993 across all classes. This high mAP indicates an exceptional balance between precision and recall, signifying the model's effectiveness in object detection. Furthermore, the model yields an impressive F1-score of 0.99 at a confidence threshold of 0.624, demonstrating a robust balance between precision and recall, which is crucial for minimizing both false positives and false negatives. This prototype being developed assumes that a destination can be set by an operator of the system using the server that connects to the CCTV camera. The system was tested in enclosed environments and was able to provide a path that potentially avoids obstacles. The development of audio commands to guide visually impaired users is ongoing. These audio commands depend on identifying the direction an individual is going, requiring an additional deep-learning model to generate accurate instructions.