Straightforward Recognition of Daily Objects in Smart Environments from Wearable Vision Sensor

Abstract

In this work, we propose a method to create and synthesize a new set of virtual images of daily objects within a smart environment partially automating the labeling process. Proposed methods enable the generation of a large dataset from a set of few images using an ad hoc data augmentation, which increases the original dataset size, generating new items through partial modification of available images. The proposed method for data augmentation is accomplished through the following steps: (i) object tracking is proposed to identify and label static objects; and (ii) background subtraction is used to select the masked foreground object of moving objects, which are virtually projected with geometry transformation over room images used as background. Furthermore, a case study is carried out, where an inhabitant wears a wearable vision sensor in a daily scene. Eight objects are learned using the proposed methodology. Finally, obtained results and successful recognition rates are discussed.