Dealing with human variability in motion based, wearable activity recognition

We describe a novel algorithm for the spotting and recognition of human activities from motion sensor signals. Our work focuses on being able to deal with the variability of human actions including user independent training. The core idea is that most actions can be divided into segments that allow a high degree of variability and segments that, due to physical constraints, have to be executed in a fairly invariant way. In the paper we present a method for identifying such segments and using them for the spotting and classification of complex activities. We evaluate our method on a well known car assembly data set and show that it performs significantly better than previous approaches.