Long-term Activities Segmentation using Viterbi Algorithm with a k-minimum-consecutive-states Constraint

Abstract

In the last years, several works have made use of acceleration sensors to recognize simple physical activities like: walking, running, sleeping, falling, etc. Many of them rely on segmenting the data into fixed time windows and computing time domain and/or frequency domain features to train a classifier. A long-term activity is composed of a collection of simple activities and may last from a few minutes to several hours (e.g., shopping, exercising, working, etc.). Since long-term activities are more complex and their duration varies greatly, generating fixed length segments is not suitable. For this type of activities the segmentation should be done dynamically. In this work we propose the use of the Viterbi algorithm on a Hidden Markov Model with the addition of a k-minimum-consecutive-states constraint to perform the long-term activity recognition and segmentation from accelerometer data. This constraint allows the algorithm to perform a more informed search by incorporating prior knowledge about the minimum duration of each long-term activity. Our experiments showed good results for the activity recognition task and it was demonstrated that the accuracy was significantly increased by adding the k-minimum-consecutive-states constraint.