Real-time Outdoor Objects Recognition and Distance Detection For Visually Impaired People

Abstract

Applications of computer vision technologies aimed for helping visually impaired people has long been an important topic in the society. According to the study, low mobility is one of the major daily life problem encountered by the visually impaired. However, majority of the existing applications on smartphone are not specialize in identifying outdoor objects, processing speed are slow due to the high latency of cloud computing. It fails to provide timely notifications and warnings regarding the objects surrounding the individual. In this project, I developed an offline smartphone application that performs real-time object recognition and distance detection on common outdoor objects with the aim of assisting visually impaired to navigate around unfamiliar locations. The system can identify dangerous objects, including ladder and bicycle, and assist users in commuting by locating bus stops and MTR entrance. Instant feedback will be given to user when the detected distance is smaller than a threshold value. The YOLO algorithm is trained with custom object dataset using Darknet neural network framework. The model achieved 67.77% recognition accuracy at 0.072 false positive per image. The first demonstration has been conducted and feedback has been collected at Hong Kong Blind Union.