Activity Recognition in Assisted Living Facilities with Incremental, Approximate Ground Truth

In this paper we present the problems associated with acquisition of ground truth, which is a critical step in facilitating accurate and automated care in Assisted Living Facilities. The approach permits both bottom up and top down methods of reasoning about data. The tradeoffs between granularity of ground truth acquisition and its impact on the detection rate are presented. It is suggested that the acquisition of ground truth should become a seamless operation incorporated transparently into the workflow of operations in these facilities. It is expected that with automation of collection, the increasing corpus of ground truth will lead to steady improvements in the detection rate and therefore the quality of automated monitoring and care provisioning. The methodology and models are substantiated with real data from two assisted living facilities, one in Singapore and the other in France. Although the results are preliminary they are quite promising.