Temporal Segmentation using Support Vector Machines in the context of Human Activity Recognition

Abstract

In this thesis we construct a method to create a temporal segmentation on time series data recorded during human activities by inertial sensors in smartphones. Our assump- tion is that information about changes between activities can aid activity classification methods, since it adds information about sequential data points. We explore current temporal segmentation methods and relate the problem to change detection. A change in the performed activity results in a change in the data distribution of a time window. Using a One-Class Support Vector Machine algorithm we model the data distribution in the shape of a high dimensional hypersphere. A change in the distribution can then be discovered by analyzing the radius of the hypersphere, which will increase in such events.

We implement an algorithm to process real-world time series data and find changes in performed activities. The data needs to be recorded in a continuous nature, with the transition periods between activities present. Since common available data sets do not meet this requirement, we have recorded our own data set and made it publicly available. A ground truth is provided both as manually annotated change points and video recordings of the activities.

The constructed method, One-Class Support Vector Machine (SVM) Human Activity Temporal Segmentation, is able to find changes between performed activities in an un- known environment.