Towards Outdoor Navigation System for Visually Impaired People using YOLOv5

Navigation is one of the many difficulties the Blind and Visually Impaired (BVI) face, be it indoor or outdoor. BVI people use their remaining senses such as smell, touch, and memory to navigate in a familiar environment. But the real problem begins when the environment is new. In the new environment, they seek help from closed ones or use assistive tools. This paper proposes a deep learning based outdoor smartphone navigation system approach which will assist the user in crossing the road by detecting the crosswalks from a mobile device. CNN of Deep learning has been proven to give the best results for image classification. Hence we want to analyze and compare different variants of CNN. As mobile devices have computed limitations, we explore two lightweight state-of-the-art neural networks, namely YOLOv5 and MobileNetSSDV2, for this task. The image dataset of crosswalks for work is collected via web scraping and manual effort. To evaluate the performance of the two neural networks: mAP, loss, and training time of the models are compared. As a result, it is determined that YOLOv5 has a five times faster inference speed than MobileNetSSDv2, which makes YOLOv5 more suitable for real-time applications. This approach utilizes the smartphone camera to detect the crosswalks and provide voice feedback to help people with low vision understand their environment.