%Aigaion2 BibTeX export from IfE Document Management System
%Monday 21 May 2012 09:05:41 AM

@INPROCEEDINGS{Calatroni11,
     author = {Calatroni, Alberto and Roggen, Daniel and Tr{\"{o}}ster, Gerhard},
   keywords = {OPPORTUNITY},
      month = jun,
      title = {Automatic Transfer of Activity Recognition Capabilities between Body-Worn Motion Sensors: Training Newcomers to Recognize Locomotion},
  booktitle = {Eighth International Conference on Networked Sensing Systems (INSS'11)},
       year = {2011},
    address = {Penghu, Taiwan},
   abstract = {Wearable human activity recognition systems rely on one or more body-worn motion sensors. Those are becoming increasingly popular, as we find them in most mobile phones and they start to appear in smart shoes and garments. One of the weaknesses of current systems is that any sensor node must be trained to recognize activities in a time-consuming process requiring explicit user intervention. In a pervasive computing scenario, users can buy new "general purpose", untrained, sensor-enabled devices to upgrade or replace their on-body sensing infrastructure. Yet they want to keep the recognition capability of the previous system. In this paper we address the problem of transferring, without user intervention, the activity recognition capability of an existing sensor node to the newly deployed and untrained node. We present a sensor-modality-independent method assuming the co-presence of trained sensor nodes and untrained nodes, to operate an autonomous transfer of the activity recognition capabilities. Labels of recognized activities from the first nodes are transferred to the latter, which incrementally associate their sensor signals to the received labels. We compare this method with a naive approach where the activity models are directly transferred between the nodes, and we use manual training of the nodes as a common baseline. We assess the performance of the methods on a realistic dataset with eight body-worn sensors in a task of posture and modes of locomotion recognition. With the novel method, a newly deployed sensor reaches recognition accuracy in average within 9.3\% of the baseline and 17.3\% higher than the naive approach. In addition, the novel method allows transfer between nodes of different modalities and placement.}
}