Development of a Wearable Household Objects Finder and Localizer Device Using CNNs on Raspberry Pi 3

Abstract

Aging society is the future for most of countries and it creates many issues for both elderly and other family members. Working age members have to work outside in the day time to get enough income for the family. Elderly are forced to live alone and facing a difficulty because they cannot see things clearly. In this research, Convolutional Neural Networks (CNNs) for household objects finder and localizer is proposed. Raspberry Pi 3 is used to create a wearable device for elderly to help see objects and tell the type and location of the objects via voice. YOLO neuron network is selected because of the speed and resource consumption. Data set of household objects images was prepared and fed into YOLO network as training inputs. Training process is done on a desktop computer before transferred into Raspberry Pi 3. Trained YOLO can classify and local several objects in the same time with accuracy of 83%. By using PeachPy and NEON instruction set for compiling YOLO in ARM environment, the speed of detection and localization in Raspberry Pi 3 equals to 0.8 seconds per image. Prototype of wearable device is built by using a camera, headphone, push switch, battery and Raspberry Pi 3. Device will capture an image when a push switch is pressed and Raspberry Pi returns pre-recorded voice results consist of objects types and their locations through headphone to elderly.