Deliverable 4.5 Evaluation of the activity recognition system

This deliverable reports the evaluation of the activity recognition system in household chores in WP4 of the ACCOMPANY project.
We collected the ACCOMPANY dataset in Troyes. The dataset contains 6 subjects performing daily household activities in a nursing apartment, including 3 elderly people and 3 young people. A variety of sensors are mounted in the department, including robot camera sensors, overhead cameras, pressure sensors and contact sensors. The data is annotated by 7 independent annotators. The annotation from multiple annotators can be used as confidence of labeling, and this can facilitate robust learning of human activities.
To recognize human activities using these sensors, we have built a system to recognize lowlevel sub-activity sequences (accepted at ICRA 14') as well as a hierarchical approach for recognizing high-level activities (accepted to ROMAN 14'). Our experiments consist of multiple activities of users.
In order to incorporate confidence of annotation into our activity recognition framework, we proposed the method “soft labeling”, which allows annotators to assign multiple, weighted, labels to data segments. The work has been presented at the conference RSS 14’.