Wearable sensor activity analysis using semi-Markov models with a grammar

Abstract

Detailed monitoring of training sessions of elite athletes is an important component of their training. In this paper we describe an application that performs a precise segmentation and labeling of swimming sessions. This allows a comprehensive break-down of the training session, including lap times, detailed statistics of strokes, and turns. To this end we use semi-Markov models (SMM), a formalism for labeling and segmenting sequential data, trained in a max-margin setting. To reduce the computational complexity of the task and at the same time enforce sensible output, we introduce a grammar into the SMM framework. Using the trained model on test swimming sessions of different swimmers provides highly accurate segmentation as well as perfect labeling of individual segments. The results are significantly better than those achieved by discriminative hidden Markov models.

Inertial Sensor Technology for Elite Swimming Performance Analysis: A Systematic Review

“…Researchers have exploited these characteristics to develop methods which may be used to automatically detect the stroke type completed for any given lap [ 11 , 46 , 59 , 60 ]. Davey and colleagues [ 11 , 19 ] developed an algorithm that calculates sensor orientation and signal energy ( Figure 10 ).…”

“…Many of the systems described are prototype systems that have been developed specifically for use in swimming research [ 53 , 74 , 83 ]. Additionally, various commercially available sensor devices such as Physilog (BioAGM, Switzerland) [ 64 ]; FreeSense (Sensorize, Italy) [ 47 ]; Minimax X (Catapult Sports, Australia) [ 46 , 86 ] and Shimmer (Shimmer, Ireland) [ 78 , 94 ] have also been used. These platforms are not specifically designed for use in swimming, therefore various modifications to make them suitable for use in aquatic environments have been developed, specifically to provide waterproofing solutions.…”

Trends Supporting the In-Field Use of Wearable Inertial Sensors for Sport Performance Evaluation: A Systematic Review

“…Activity classification has been performed by measuring body segments in relative or absolute orientation, to detect sprinting, jogging, walking [ 52 , 195 , 232 , 326 ], jumping, cutting, kicking [ 52 ], dribbling, ball strike, tackling [ 232 ], swimming style [ 164 , 304 ], trampoline routines [ 161 ], climbing phases [ 114 ], rugby collisions [ 176 ], volleyball actions [ 306 ], tennis strokes [ 301 ], golf putting [ 171 ], and aerial manoeuvres in half-pipe snowboarding [ 159 ] and skateboarding [ 154 ].…”

“…Supervised approaches train the classification system using a set of data with a known classification. The classifiers of this kind used in the revised literature are decision trees (random forest training algorithms [ 52 , 326 ], logistic model trees [ 232 , 326 ]); stepwise discriminant analysis of principal components [ 184 ]; k-Nearest Neighbour algorithms (IBk [ 232 ], Lazy IBk [ 52 ]); Artificial Neural Networks (RBF Network [ 52 ], MLP network [ 232 ]); Support Vector Machines [ 176 , 232 , 326 ]; Naïve Bayesian classifiers [ 52 , 154 , 232 ]; fractals [ 94 ]; Hidden Markov or Semi-Markov models [ 171 , 304 ]; and Hidden Conditional Random Field [ 176 ].…”

Smart Annotation of Cyclic Data Using Hierarchical Hidden Markov Models

“…Thomas et al (2010) investigated semi-supervised SMMs for simultaneous segmentation and classification. However, their method requires some fully-labeled data, as well as prior expert knowledge of the likely sequence of events [ 20 ]. The rhythmic extended Kalman filter has also been used for unsupervised gait analysis; however, it relies on a body model, assumes walking in a straight line and requires multiple sensors firmly attached to specific locations on the subject, so although it is online and unsupervised, it requires patient-specific information and may not be extensible to other cyclic movements.…”