Pedestrian lane and obstacle detection for blind people

Abstract

Pedestrian lane and obstacle detection have been a big problem for a blind person. The person has to use special tool or assistant to do daily activities. This work has utilized camera function, connected with data processing (mini PC) to process information captured by a camera. At first, RGB image, which was captured camera, was converted into XYZ colour system. This colour system was very useful to highlight pedestrian lane to reduce other objects. Then, colour filter was implemented to remove unnecessary objects, followed by close morphology to expose pedestrian lane. The result is white region which represent pedestrian lane. Major axis was then calculated using moments and its angle (calculated counter clockwise with reference to x-axis) was sent to the user to notify him/her which direction he/she can go. In the process of obstacle detection, some samples of RGB images has been used to train a neural network. The model was then used as obstacle detector. RGB images captured by camera were then used as test data. The result > 0.7 was considered as obstacle. The experiment shows that under illumination of <15000 lux, the method can achieve 89.7 percentage accuracy on pedestrian lane detection and 100 percentage accuracy on obstacle detection.