Journal of Biomechanics 41 (2008) 216–220 Short communication A hidden Markov model-based stride segmentation technique applied to equine inertial sensor trunk movement data Thilo Pfau à , Marta Ferrari, Kevin Parsons, Alan Wilson Structure and Motion Laboratory, Department of Veterinary Basic Sciences, The Royal Veterinary College, University of London, Hawkshead Lane, North Mymms, Hatfield AL9 7TA, UK Accepted 14 August 2007 Abstract Inertial sensors are now sufficiently small and lightweight to be used for the collection of large datasets of both humans and animals. However, processing of these large datasets requires a certain degree of automation to achieve realistic workloads. Hidden Markov models (HMMs) are widely used stochastic pattern recognition tools and enable classification of non-stationary data. Here we apply HMMs to identify and segment into strides, data collected from a trunk-mounted six degrees of freedom inertial sensor in galloping Thoroughbred racehorses. A data set comprising mixed gait sequences from seven horses was subdivided into training, cross-validation and independent test set. Manual gallop stride segmentations were created and used for training as well as for evaluating cross-validation and test set performance. On the test set, 91% of the strides were accurately detected to lie within 740 ms (o10% stride time) of the manually segmented stride starts. While the automated system did not miss any of the strides, it identified additional gallop strides at the beginning of the trials. In the light of increasing use of inertial sensors for ambulatory measurements in clinical settings, automated processing techniques will be required for efficient data processing to enable instantaneous decision making from large amounts of data. In this context, automation is essential to gain optimal benefits from the potentially increased statistical power associated with large numbers of strides that can be collected in a relatively short period of time. We propose the use of HMM-based classifiers since they are easy to implement. In the present study, consistent results across cross-validation and test set were achieved with limited training data. r 2007 Elsevier Ltd. All rights reserved. Keywords: Pattern recognition; Hidden Markov model; Inertial sensor; Stride segmentation; Racehorse 1. Introduction Three degrees of freedom (3DoF) inertial orientation sensors are now of sufficiently small size and weight to be used for kinematic measurements in humans (Goodvin et al., 2006; Luinge and Veltink, 2005; Zhou et al., 2006, 2007). In addition to accurate orientation data they can be used to calculate high accuracy displacement data (Pfau et al., 2005). This technique has been applied for measure- ment of linear and rotational kinematic variables on animals during over ground locomotion (Pfau et al., 2006, 2007; Parsons et al., under review). Compared to traditional gold standard kinetic (force platforms) and kinematic (e.g. 3D optical motion capture) data collection techniques, inertial sensors can provide continuous data from unrestricted movement when sensors are mounted on the subject. In combination with telemetry or data loggers, several hours of data can be collected at a time (e.g. Pfau et al., 2006). The large data volume enables novel statistical approaches to data analysis but a significant degree of automation is required in the processing to prepare the data for these analyses whilst maintaining realistic work- loads. In addition, these data will typically include periods of different activities, e.g. different gaits and speeds in animals and sitting/driving/walking in people. After the collection, classification into gait categories is required and individual strides have to be identified and segmented for subsequent stride averaging or grouping into speed categories (e.g. from a GPS unit, Pfau et al., 2006). ARTICLE IN PRESS www.elsevier.com/locate/jbiomech www.JBiomech.com 0021-9290/$ - see front matter r 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.jbiomech.2007.08.004 à Corresponding author. Tel.: +44 1707 666327. E-mail address: tpfau@rvc.ac.uk (T. Pfau).