Development of Prototype for Obstacle Detection for Visually Impaired using Defocus Measure

Abstract

The major challenges faced by visually impaired are mobility and orientation during their navigation.  For safe navigation of visually impaired, an assistive system that can estimate the distance between visually impaired and the obstacle (pedestrian) and generate an alert signal for the user is required. This paper presents a prototype of obstacle (pedestrian) distance estimation and alert generation system using a defocus measure for visually impaired.  Initially, the video frames are extracted and preprocessed. Further, the feature points are detected using the Shi Tomasi corner detector. The feature points in the subsequent frames are tracked using Kanade-Lucas-Tomasi (KLT). Region of interest of size 20 x 20 pixels is extracted around every feature point and the defocus measure is computed by using first level wavelet decomposition. For a typical camera, the defocus measure is a function of the distance between camera and obstacle and therefore can be used to estimate distance. The algorithm is implemented on the ARM Cortex-A7 platform using Embedded Linux and open-source computer vision library. The algorithm is tested on a USB web camera and 10 videos. This method is proved to be effective forobstacle distance measurement. The average absolute error of 0.0081m with a standard deviation of 0.02775 and the average error of 0.347% with a standard deviation of .63583 is observed for camera 1. The average absolute error of -0.21m with a standard deviation of 0.359 and the average error of -4.69% with a standard deviation of 6.99is observed for camera 2.The alert signal is generated at a distance of 3.33m for visually impaired during the navigation.The repeatability of the prototype is 90%.