Smart visual assistant

Abstract

An extendible object recognition platform has been created that employs the use of several feature detection algorithms and feature descriptors to accurately match keypoints of images. The platform is applied to an application which performs generic recognition between sets of images. The application stores a database of images tagged with names of objects that they represent. More images can be added to the database through the application. Object recognition is accomplished by matching frames from a video-stream with the database of tagged images and using a match validation technique which filters the results of a k-NN search based on the distance of the two nearest neighbours of a query point. The application performs all processing on the device in real-time and allows the user to switch between different feature detection algorithms and descriptors on the fly. Each approach was evaluated in a series of tests which measures performance in terms of efficiency and accuracy. SURF was found to be the best overall performer giving accurate results in real-time and scaling well through the use of a fast approximate k-NN search.