Object Recognition to Support Navigation Systems for Blind in Uncontrolled Environments

Abstract

Efficient navigation is a challenge for visually impaired people. Several technologies combine sensors, cameras, or feedback chan-nels to increase the autonomy and mobility of visually impaired people. Still, many existing systems are expensive and complexto a blind person’s needs. This work presents a dataset for indoornavigation purposes with annotated ground-truth representingreal-world situations. We also performed a study on the efficiencyof deep-learning-based approaches on such dataset. These resultsrepresent initial efforts to develop a real-time navigation systemfor visually impaired people in uncontrolled indoor environments.We analyzed the use of video-based object recognition algorithms for the automatic detection of five groups of objects: i) fire extin-guisher; ii) emergency sign; iii) attention sign; iv) internal sign, and v) other. We produced an experimental database with 20 minutesand 6 seconds of videos recorded by a person walking throughthe corridors of the largest building on campus. In addition to thetesting database, other contributions of this work are the study onthe efficiency of five state-of-the-art deep-learning-based models(YOLO-v3, YOLO-v3 tiny, YOLO-v4, YOLO-v4 tiny, and YOLO-v4scaled), achieving results above 82% performance in uncontrolledenvironments, reaching up to 93% with YOLO-v4. It was possible toprocess between 62 and 371 Frames Per Second (FPS) concerning thespeed, being the YOLO-v4 tiny architecture, the fastest one. Codeand dataset available at: https://github.com/ICDI/navigation4blind.