Low Cost Activity Recognition Using Depth Cameras and Context Dependent Spatial Regions

Michael Karg (1) , Alexandra Kirsch (1, 2, 3)

1 Institute for Advanced Studies
2 Intelligent Autonomous Systems
3 Human-Computer Interaction and Artificial Intelligence

To be useful helpers for humans in domestic environments, robots should be aware of human task execution to anticipate and adequately react to human actions. Hence the field of activity recognition has become of increasing interest in the robotics community and many approaches are based on sequences of object detections or human posture recognition, requiring the environment to be equipped with loads of sensors or extremely expensive motion tracking systems. In this paper we investigate the use of inexpensive depth cameras to perform activity recognition using context dependent spatial regions with two different approaches for activity recognition: Spatio-Temporal Plan Descriptions and Hierarchical Hidden Markov Models. We evaluate both approaches in a simulated and a real-world environment, showing that reliable activity recognition is possible using a sensor setting for less than 250 $ in a spatially limited environment.

Domaines: Informatique [cs]; Intelligence artificielle [cs.AI]; Interface homme-machine [cs.HC]; Robotique [cs.RO]

Michael Karg, Alexandra Kirsch. Low Cost Activity Recognition Using Depth Cameras and Context Dependent Spatial Regions. 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2014), 2014, Paris, France. ⟨hal-01691662⟩