Trainable Visual Aids for Object Detection and Identification

Johns Hopkins University, Baltimore MD

Investigators

Abstract

This project leverages advances in statistical learning theory, machine vision, and massively parallel very-large-scale-integration technology to develop a custom-trainable, versatile, self-contained, and mobile system for visually impaired users. The system will aid the user in interacting freely with other people and the environment, by rapidly detecting and localizing key visual environmental cues and rapidly recognizing and identifying familiar people and objects. At the core of the system is the "Kerneltron", a massively parallel Support Vector "Machine" (SVM) in silicon. The SVM hardware will be trained on-line by the end user to accommodate a variety of visual detection and recognition tasks in everyday situations through presentation of examples. The recognition core will be embedded in a portable prototype visual aid, interfacing with a CCD camera front-end, and an audio synthesizer back-end. Menu-driven keypad control will allow direct input and feedback from the user in training and directing the system. The user interface will be based on "OpenEyes", a wearable computer vision system for the blind. Proof of concept demonstration of the hardware system and evaluation of the training and test performance will be conducted with feedback from volunteer impaired users.