Measuring Behavior 2005: Wassink and Baten

Automated dynamic activity monitoring

R. Wassink and C. Baten

Roessingh Research & Dev., Enschede, the Netherlands

In some applications of ambulatory assessment of 3D human motion typically detailed kinematic and kinetic biomechanical data is gathered or derived over longer periods (hours), e.g. in ergonomic applications. For sensible interpretation of this data continuous context information is required, e.g. ‘activity performed’. Currently this data typically is gathered by labor intensive manual observation, while all other data gathering is fully automated. To get rid of the labor intensive manual observations a new method is proposed for automated activity monitor for human activities using the data already automatically gathered in the kinematic assessment.

Also in clinical motion analysis applications of ambulatory recording technologies (portable gait lab) a need has risen for automated context classification. Typically data is gathered of many motion cycles (e.g. steps in Gait in analysis) and to facilitate wide spread use by health care professionals automated cycle based interpretation assistance is required. For e.g. gait analysis this means estimation of kinematic and kinetic step cycle parameters and typical derived statistics. A requirement for this is automated step cycle detection and step cycle classification. Both application areas of ambulatory motion analysis require automated activity recognition and classi.cation including estimation of start and end times.

Methods

The proposed method applies self learning Hidden Markov Modeling (HMM) technique to the kinematic data. In a training phase for each activity a HMM is derived from training data. In the application phase for each the HMM the probability is estimated that the current activity is the one represented by this HMM. In a post processing phase these probabilities are used to decide on which activity is currently recognized. In a more sophisticated estimator also (estimated) a priori probabilities of occurrence are taken into account in the post-processing.

Results

Pilot experiments have indicated that HMM methodology seems very capable of delivering the requires recognition and classification functionalities. Current research aims at examining accuracy, robustness and generalizibility of the methodology. In this SIG contribution possibilities and experiences with applying HMM will be discussed as well as its potential for a general Activity Monitor for use in functional evaluation in rehabilitation and ergonomy.

Paper presented at Measuring Behavior 2005 , 5^th International Conference on Methods and Techniques in Behavioral Research, 30 August - 2 September 2005, Wageningen, The Netherlands.