Embedded Vision-Based Intelligent Device for the Visually Impaired

Abstract

Computer Vision, a branch of Artificial Intelligence has been showing tremendous potential in solving day to day problems. The world has been tuning itself to the language of embedded systems. Dedicated, distributed and intelligent systems are making lives easier. The important things about intelligent embedded systems have started becoming more important. Computer Vision is achieved using deep learning, one of the most computationally expensive domains in computing. Deploying a computer vision solution in the form of deep learning models on single board computers that work with constrained resource allocation is indeed a challenge. The problem that this work takes aims to help visually impaired people by augmenting the way they interact with their surroundings. The proposed systems aim to make use of Embedded Vision-Computer Vision deployed on a single board computer like the Raspberry Pi, to recognize and convey emotions, age and gender of a person in front of a visually impaired person to him as audio output. Wide ResNet is used to implement age and gender classifier, while emotion classification uses a mini Xception Net with ImageNet weights, fine-tuned over the FER-2013 dataset. The reasons for selecting the respective algorithms and their deployment on the hardware shall be discussed. The system helps in adding more behaviour to the way a visually impaired person interacts with his or her peers. A few critical hardware deployment strategies that the work entails shall also be discussed.