Multi-Type Activity Recognition in Robot-Centric Scenarios

TL;DR

This paper introduces a unified approach for recognizing multiple activity types in robot-centric scenarios using a new descriptor, Relation History Image, enabling detection of sequential or concurrent activities.

Contribution

The paper presents the Relation History Image descriptor and an optimization method for recognizing diverse activity types in a unified framework, applicable to robot-centric data.

Findings

Effective detection of multiple activity types in robot scenarios

Superior performance of RHI descriptor on new and public datasets

Robust recognition of activities occurring sequentially or concurrently

Abstract

Activity recognition is very useful in scenarios where robots interact with, monitor or assist humans. In the past years many types of activities -- single actions, two persons interactions or ego-centric activities, to name a few -- have been analyzed. Whereas traditional methods treat such types of activities separately, an autonomous robot should be able to detect and recognize multiple types of activities to effectively fulfill its tasks. We propose a method that is intrinsically able to detect and recognize activities of different types that happen in sequence or concurrently. We present a new unified descriptor, called Relation History Image (RHI), which can be extracted from all the activity types we are interested in. We then formulate an optimization procedure to detect and recognize activities of different types. We apply our approach to a new dataset recorded from a robot-centric perspective and systematically evaluate its quality compared to multiple baselines. Finally, we show the efficacy of the RHI descriptor on publicly available datasets performing extensive comparisons.