Modelling and implementation of real time texture detection module for mobility assistant for visually impaired system (MAVI)

Abstract

As the volume of video data is increasing day by day, real-time video processing is becoming an important application to build any real time device based on image processing. An interesting task that is crucial for the safety of a person is safety from moving vehicles. This can be a crucial module for the people who are interested in developing an assisting device for visually impaired that can help in identifying the type of surface a person is walking on. The thesis presents the design space exploration, implementation of texture detection module for MAVI (Mobility Assistant for Visually Impaired), an outdoor navigation system in Indian context, for helping visually impaired people by making them aware of the surface on which they are walking on. It will help them in identifying the sidewalks (pavement) to ensure safety from moving vehicles on a road. The work involves exploring various models which will work in real time, based on di_erent texture features like LBP and GLCM and using Computer Vision techniques, measuring their accuracies and performance by cross compiling their code on ZedBoard. It also explores the usage of deep learning model like SegNet for this task and its performance in real time. The major challenge for such a module is to perform well in a lot of diversity and complexities that arise in the Indian scenario due to non standard practices, therefore its scope is limited to three major texture classes i.e., road, pavement and grass