Detection and Distance Estimation against Motorcycles as Navigation Aids for Visually-impaired People

Visually impaired people are limited to detecting obstacles when walking in an outdoor environment. This research aims to detect, estimate distance and relative position of the visually impaired people to the obstacle, particularly parked motorcycle. The proposed system uses the Single Shot Multibox Detector (SSD) Mobilenet for parked motorcycle detection by testing several different learning rates. The Pinhole Camera Model method is used to estimate the distances from 2 to 5 meters by comparing the actual motorcycle with the motorcycle in the image based on the similar triangle principle. The last stage is estimating the relative normal horizontal position of the visually impaired people to the obstacle to determine whether the position of the visually impaired people is in line with the right, the middle, or the left side of the parked motorcycle. This research uses 500 images, 80% as training data, and 20% as testing data. There are four scenarios based on the user's height with six parked motorcycles positions, i.e., the front, rear, right, left, right oblique, and left oblique. In addition, the tilting position of the smartphone camera uses two different angles, depending on the user's height. The highest parked motorcycle detection results obtained at the learning rate 10
^-3
with an accuracy of 95.04%, the average RMSE of the estimated distance is 0.11, and the highest accuracy of the estimated relative position is 83.26% at 2 meters distance.