Advanced Obstacle Detection Based on YOLOv5 for Safer Navigation in Autonomous Smart Wheelchair

Wheelchair users commonly rely on the help of others to ambulate. The lack of supervision from others can increase the risk of accidents for wheelchair users, especially in environments where wheelchair accessibility is limited. One of the main challenges for wheelchair users is encountering road obstacles such as stairs, uphill road, and downhill road. To address this issue, the author proposes an architectural obstacle detection system utilizing cameras to support the autonomous wheelchair system, thereby enhancing user safety. The way the system works is it will stops when stair ascents or descents are detected, increases speed when uphill road are detected, and decreases speed when downhill road are detected. The method employed for obstacle detection involves utilizing YOLOv5, a highly efficient and fast deep learning-based object detection model. The main advantages of YOLOv5 include its high speed in real-time object detection and its ability to achieve good accuracy even with small or hard-to-see objects. Three YOLOv5 model types named YOLOv5n, YOLOv5s, and YOLOv5m are compared in terms of accuracy and computational speed using NVIDIA Jetson TX2 device that embedded in smart wheelchair. This research collected 3400 architectural obstacle images from the environment of the Faculty of Computer Science, Brawijaya University, as the dataset. The results of this study show that YOLOv5n has the faster computational time, approximately 0.0729 seconds with an accuracy of 86.38%, while YOLOv5m has the highest accuracy 93.47% with computation time of 0,23342 seconds.