Discriminative Temporal Smoothing for Activity Recognition from Wearable Sensors Jaakko Suutala, Susanna Pirttikangas, and Juha R¨ oning Intelligent Systems Group, Infotech Oulu Computer Engineering Laboratory P.O. Box 4500, FIN-90014 University of Oulu, Finland E-mail: jaska, msp, jjr @ee.oulu.fi Abstract. This paper describes daily life activity recognition using wearable ac- celeration sensors attached to four different parts of the human body. The experi- mental data set consisted of signals recorded from 13 different subjects perform- ing 17 daily activities. Furthermore, to attain more general activities, some of the most specific classes were combined for a total of 9 different activities. Simple time domain features were calculated from each sensor device. For the recogni- tion task, we propose a novel sequential learning method that combines discrim- inative learning of individual input-output mappings using support vector ma- chines (SVM) with generative learning to smooth temporal time-dependent activ- ity sequences with a trained hidden Markov model (HMM) type transition prob- ability matrix. The experiments show that the accuracy of the proposed method is superior to various conventional discriminative and generative methods alone, and it achieved a total recognition rate of 94% and 96% studying 17 and 9 differ- ent daily activities, respectively. 1 Introduction Activity recognition from wearable sensors has become an important research topic in recent years [1], [2], [3]. Successful recognition of basic human activities based on sensing of body posture and motion can be used in different applications, such as health care, child care and elderly care, as well as in personal witness monitoring. In addi- tion, it provides a mechanism for using the activities to control devices around us, for example to provide personalized services to assist those with physical disabilities or cognitive disorders. In this paper we present a novel method for activity recognition from wearable sensors. It combines ideas from two major categories of supervised machine learning: discriminative and generative learning. Discriminative learning (e.g., kernel methods [4], [5], [6]) provides an effective framework for learning direct input-output mapping from a labeled training data set ( and , where presents :th input feature vector and is :th target class label) for particular appli- cations, such as classification and regression to predict unknown examples. However, adapting a discriminative framework to more advanced learning problems, such as cases where input or/and output spaces can have a structure (a sequence, for example), is not straightforward.