A Dataset of Naturally Occurring, Whole-Body Background Activity to Reduce Gesture Conflicts

TL;DR

This paper introduces a methodology and dataset for understanding natural body movements to improve gesture recognition systems, reducing false positives caused by background activity in real-world settings.

Contribution

It provides a reusable methodology and software tools to generate standardized background activity datasets for whole-body input systems.

Findings

Created a background activity dataset for a living room setting.

Demonstrated how the dataset can help redesign gesture vocabularies.

Made tools and dataset publicly available.

Abstract

In real settings, natural body movements can be erroneously recognized by whole-body input systems as explicit input actions. We call body activity not intended as input actions "background activity." We argue that understanding background activity is crucial to the success of always-available whole-body input in the real world. To operationalize this argument, we contribute a reusable study methodology and software tools to generate standardized background activity datasets composed of data from multiple Kinect cameras, a Vicon tracker, and two high-definition video cameras. Using our methodology, we create an example background activity dataset for a television-oriented living room setting. We use this dataset to demonstrate how it can be used to redesign a gestural interaction vocabulary to minimize conflicts with the real world. The software tools and initial living room dataset are publicly available (http://www.dgp.toronto.edu/~dustin/backgroundactivity/).