Journal Logo

SPECIAL COMMUNICATIONS: Methodological Advances

Activity Mapping of Children in Play Using Multivariate Analysis of Movement Events


Author Information
Medicine & Science in Sports & Exercise: January 2020 - Volume 52 - Issue 1 - p 259–266
doi: 10.1249/MSS.0000000000002119


Physical inactivity is one of the major causes of death worldwide (1) and so exercise and activity programs, designed to avoid sedentary lifestyles, are increasingly prevalent and have been shown to reduce risk factors, such as type 2 diabetes, heart disease, and even some cancers (2). For these reasons, it is important for children to develop the healthy habit of frequent physical activity (3,4) and considering that children spend a large amount of their day at school, recess becomes a natural time to encourage this. A wide range of factors determine children’s propensity for activity (5); however, it is clear that good playground design can have a positive effect (6,7). There is also a growing realization that the quality of activity, as reflected in the movement competence of individuals is as important as the quantity of exercise undertaken (8,9).

Given this importance of exercise to a healthy lifestyle, detailed and quantitative assessment of activity frequency and intensity is a well-established research area (10,11). Measurement is often by wearable, inertial sensors (i.e., accelerometers) (12,13), placed at various locations on the body (14), to give signals that are proportional to the intensity and direction (magnetometer) of movement (15). Although the implementation of this technology to obtain a faithful record of body acceleration is relatively straightforward, the interpretation of the data to inform on activity is more difficult. In particular, the wide range of movements and irregular intensity of activity displayed by children during free play (16) present a demanding challenge to quantitative analysis. The commonly used metric for assessment of activity is “counts”—integrated acceleration-magnitude during a defined epoch (17). This gives a ready measure that directly correlates to energy expenditure; however, it conveys no information on the form of movement undertaken and by definition is insensitive to rapid changes in activity level. We, therefore, present a technique, based on the raw acceleration trace and the use of a standardized reference signal—a “movement motif,” which is a data sequence corresponding to a known motion, such as a walk step or run stride. The movement motif provides a time series template (18,19) to which movement events within the acceleration trace can be compared, using pattern matching implemented with a dynamic time warping algorithm (20). The evaluation of activity through analysis of specific movement features and patterns is well established, with many variations of machine learning approaches reported (21). The aim in these studies is to classify activity into identifiable phenotypes such as walking, running, standing, and so on (22,23). Our approach differs in that although a reference motif is classified (a walking stride), individual movement events are not. We compare rather than classify. This removes any restriction on the form of movement, thus avoiding misclassification errors, while maintaining the context that a known movement type provides. Each comparison of a movement event with the motif quantifies the intensity and duration of the movement relative to the reference point of the motif sequence. This provides quantitative assessment of both quantity and form of motion undertaken during an activity session. The definition of specific movement categories for traditional pattern classification necessitates the use of extended time sequences (multiple strides) to ensure that example sets are uniform enough to describe a single class (17,24–26). In the approach presented here, there is no such restriction, and event-motif comparisons are made with accelerometer time–sequence data acquired at 40 Hz. Implementation of the technique on children’s motion data obtained from a 40-min school play period, allows high resolution temporal profiling of their activity (27,28). Participant profiling, based on the multiparameter movement metrics, is presented and used to assess variation within the cohort (29) and day-to-day trends across a week of measurements.


Participants and settings

The study was based on a set of 24 children, whose motion was recorded for one school-week (5 d), in a primary school in the United Kingdom, two children were absent on one of the days, and so, the total data set included 118 motion records. In the participant sample set, 12 of the children were in year 5 and 12 in year 6, 18 children were boys and the summary statistics of the cohort are as follows: age, 10.5 ± 0.6 yr; height, 1.44 ± 0.09 m; mass, 39.6 ± 9.5 kg; body mass index (BMI), 18.8 ± 3.1 kg·m−2. The participants wore ankle-mounted accelerometers during school recess for 5 d. The participants’ BMI, height, weight, sex, and school year were registered, and the distributions of these metrics were typical for the age group of children. A stadiometer (Holtain, Crymych, UK) and digital scales (SECA, Hamburg, Germany) were used to measure stature (to the nearest 0.01 m) and BMI (to the nearest 0.1 kg), respectively, following standard procedures. Furthermore, children were classified as either underweight (<5th percentile) (n = 1), normal weight (5th to 85th percentile) (n = 16), overweight (>85th to <95th percentile) (n = 5) or obese (≥95th percentile) (n = 2). (For more information about the participants, see Supplemental Digital Content 1, Appendix—supplementary information, The data were recorded with consent from the legal guardians and assent from the children, following the guidelines and policies of the institutional ethics committee and the Declaration of Helsinki.


The children’s motion was evaluated during normal time school-time recess (40 ± 4 min·d−1) for 5 d. A custom Micro Electro-Mechanical System based device was used to measure their physical activity at a frequency of 40 Hz and record it onto a microSD card (30). The sensor system incorporated a triaxial accelerometer with a ±16g dynamic range, 3.9 mg point resolution (with an amplitude coefficient of variation of 0.004 at 40 Hz) (ADXL345 sensor, Analog Devices). It was housed in a small plastic case and affixed via a Velcro strap to the lateral malleolar prominence of the fibula of the right leg (see additional images, Supplemental Digital Content 1, Appendix – supplementary information,

Data extraction and analysis

All data handling and analysis were done in the Matlab 2016b environment. The total duration of play varied between 42 and 50 min, only the first 40 min of activity were analyzed, this ensured that all traces studied were of the same duration. The methods described in this section were applied to all the children’s measurements along the 5 d, unless stated otherwise. Data acquired in the radial acceleration axis were selected for analysis because this had proven to be highly informative in previous work (30), with information being contained on push-off impulse, force of heel and toe impact, and angle of leg lift. The raw acceleration time signal, with no filtering or smoothing applied, was used in all analyses. The extraction of movement metrics is based on the use of a “movement motif.” This is a short, 1.1 s accelerometer sequence from a single stride, taken by a 27-yr-old man walking on a treadmill at a speed of 4 km·h−1, with the same sensor system and attachment as used in the children’s play study. This motif sequence, of a known and well understood biomechanical movement, provides a standard reference to which all of the children’s movements can be compared. This choice of movement motif was based on a requirement for a known and well-understood motion pattern that was distinct from those of the children to avoid biasing of any comparative analyses. We need a reference that is known and unchanging. This is difficult to obtain from a child because there is high variability due to the different states of physical maturity within the chosen age group. Also, the reference comes from outside of the group, and so, we get a comparison to an independent reference rather than self-referencing within the study cohort. Selected sequences of the acceleration signal, corresponding to movement events within each child’s trace, were extracted using a threshold demarcation of 1.5g. Comparison of each movement event with the motif was done using the Matlab dynamic time warping algorithm—dtw. This provides metrics on time difference (Δt) and amplitude difference (Δd). A full mathematical description of Δt and Δd is given in the results section. Dimensional reduction of 80 metrics obtained from the time dependence of Δd and clustering of the 24 children into similar groups was done using Matlab functions for principal component analysis—pca—and dendrogram clustering—dendrogram.


Signal processing and data extraction

An example of a raw acceleration trace is shown in Figure 1. This is typical for a child at play, displaying variable and interrupted movement across the play session and complex acceleration features at short timescales, with no discernible regularity. Commonly used analyses of acceleration data take a time-averaged approach, defining “counts” and categorizing activity level by the use of signal cut-points (31). In implementing time integration, some knowledge of the form of the movement is inevitably lost. To avoid this, we take an alternative approach and implement an event-based analysis that highlights the temporal shape of the short, often subsecond, acceleration features. This provides metrics that can inform on the type of movement undertaken, with a time resolution that is consistent with biomechanical and musculoskeletal control dynamics. The challenge in doing this for children’s play data is to find a robust method for defining motion events within nonuniform acceleration traces. Our solution is to use a movement “motif”—a well-understood, standard motion pattern. This sets a reference to which the acceleration signal can be compared, and events identified as data sequences with similar amplitude and duration as the motif. Essentially, the raw acceleration trace is sectioned into events through a loose pattern matching to the motif standard. The movement motif is shown in Figure 2A and is the acceleration trace from a single stride taken by an adult (male, age 27 yr) on a treadmill walking at a speed of 4 km·h−1. Motion events are located using a peak detect algorithm with an imposed peak threshold of 1.5g and a minimum peak-to-peak distance of 40 samples (1 s). Thus, a series of short sections within the signal trace are identified, within which the acceleration is similar to or greater than that imposed when walking. The identification process also ensures that no two events temporally overlap. A short trace section with five events identified is shown in Figure 2B.

A, Typical radial acceleration trace from a child wearing an ankle-mounted sensor for a 40-min play session. B, Expanded view of a 5-s play sequence.
A, Radial acceleration trace of the walk-step motif. B, Short sample of typical acceleration trace showing event peaks detected, the motif signal is compared with each of these using DTW. C, DTW traces for a motif-event comparison with instantaneous time warps, δti, and acceleration amplitude difference, δdi. D, Scatter plot of all events within a single 40-min playground session, Δd—magnitude of fractional event-to-motif acceleration difference, Δt—fractional extension of signal due to time-warping. Red areas indicate parameter space occupied by reference data obtained from participant walking and running on a treadmill with a unit incremented speed of 3 to 13 km·h−1. DTW, dynamic time warping.

Once motion events have been identified a secondary challenge arises as to how these are to be parameterized? As Figure 1B shows, they vary considerably in shape, duration, and magnitude, thus, it is difficult to capture this heterogeneity with a tractable number of consistent metrics that can be easily extracted. We, therefore, choose to characterize each event by comparison to the motif rather than by direct measurement of the event acceleration values. Each motion event is assessed by asking the question—“how close is it to a walk step?” This produces correlation metrics that are quantitative, robust, and which provide context to the movement undertaken. The event to motif correlation is done using a dynamic time warping algorithm (32). The dynamic time warping between event and motif signals introduces time steps in the data sequence to achieve optimum matching between traces (33,34). Basically, the two signals are stretched at various time points to create “warped” sequences, these stretches are imposed in a way that achieves the best match between the pair of traces. Two parameters are extracted from each of these event-motif comparisons—the fractional change in time (Δt) and the magnitude of the acceleration difference between the time warped signals (Δd), measured as the mean per sample point. These result from summation over the full trace and are mathematically defined as:

where δti and δdi are the time and amplitude differences, respectively (see Fig. 2C), tmotif is the duration of the motif signal, and nDTW is the total number of samples in the time warped signals. These event parameters can be displayed for a complete activity session in the form of a simple scatter plot. The plot obtained from the sample trace in Figure 1 is shown in Figure 2D. This provides an individualized, contextual map of movement, and a ready visualization of the child’s activity during play. Each point identifies a movement event, thus their density quantifies the amount of activity undertaken and corresponds to the information gathered in a traditional approach, where activity counts are recorded. Here, however, there is also information on the form of each movement, captured in the x and y coordinate values. The Δd value gives an immediate indication as to the intensity of the movement with the zero point being the reference level of walking (~3 MET·h for the 4 km·h−1 motif (35)). The Δt value informs on how close the time phasing of acceleration is to a walk step. Values of Δt > 0 indicate the magnitude of the fractional difference in duration of each movement event to that of the walking stride motif. This is an absolute number and so does not differentiate between shorter or longer duration. Comparison to the motif standard also allows benchmarking of the child’s activity to that undertaken in a controlled environment. The areas outlined in red in Figure 2B indicate the range of values obtained when the motif is compared with other events in the treadmill-study acceleration signal, from which it is extracted. This shows the evolution from walk areas (low Δt, 3–5 km·h−1) to running (high Δt, 9–13 km·h−1). The red shaded area centered at zero Δd is the parameter state space covered by multiple step events acquired at 4 km·h−1 (i.e., stride-to-stride variability in the motif itself). The overlay of data from a staged walk to run exercise of the treadmill (red sections) onto the map of the child’s movement (black circles) provides an immediate visual assessment of their activity in the 40 min of play. The spread of data shows the range of activity intensity and indicates the degree of variability compared with the controlled movement on a treadmill running through a sequence of set speeds. A number of features are evident in the child’s movement map: (i) a majority of low-intensity events have greater Δt values than the motif sequence (3–5 km·h−1 section), this reflects the shorter stride duration relative to the adult walk motif; (ii) there is a wide range of movement profile with events (black circles) spanning a continuous area that encompasses the 3- to 13-km·h−1 treadmill reference set; (iii) there are ultrahigh intensity outlier events, which reflect movement for which the sum acceleration is well above that within a gait step produced by an adult running at high speed (13 km·h−1).

Multivariate profiling

The extraction of multiple parameters for profiling of children from their activity profile was based on the time dependent values of Δd (Fig. 3A) as this gave a much greater discrimination that the Δt metric. The total number of events plus the summed value of the positive and negative Δd metric within a sliding, 2-min time window provided 80 measures per child over the 40-min activity sequence. Dimensional reduction was implemented using PCA; the two-dimensional plot for all 118 activity traces is shown in Figure 3B. To interpret the PCA plot three regions were identified: L—low axis 1 and 2; M—medium axis 1, high axis 2; H—high axis 1, low axis 2 (Fig. 4A). Representative plots of the acceleration trace (magnitude) from each of these regions, are shown in Figure 4B–G. Inspection of these shows that component 1 of the PCA correlates to activity intensity, measured as mean acceleration over the duration of the activity session (Pearson, r = 0.54), (for correlation plot, see Supplemental Digital Content 1, Appendix—supplementary information,, whereas component 2 reflects differences in the time-staging of activity during play. Closely located points in the PCA plot indicate children with highly similar motion variables.

A and B, Time-dependent measures for +′ve Δd (A) and −′ve Δd (B); cumulative movement-event count—blue circles, ∑Δd within 2-min time window—black bars. C, Motion-metrics PCA for all children, the day of activity is indicated by the color shading (Monday—red, Tuesday—green, Wednesday—blue, Thursday—black, Friday—magenta).
A, Motion-metrics PCA for all children with highlighted areas. B–G, Typical acceleration magnitude traces for children within the highlighted areas of the PCA plot—L (B & E), M (C and F), and H (D and G).

Figure 5A shows an expanded view of the PCA plot (shaded area in Fig. 4A). The raw traces, from two children juxtaposed in the PCA plot, indeed confirm that their acceleration profiles are highly correlated across the whole of the play duration (Fig. 5C and D). Dendrogram plots provide an alternative to PCA for identifying hierarchical clustering of children based on their activity profiles. Figure 5B shows the dendrogram for all traces, sorted into 30 clusters using a weighted method with arithmetic mean (WPGMA), operating on the pairwise distance matrix between all points in the multivariate space (Euclidean distance). This provides information on groups of children with similar activity, for example, cluster 17 encompasses the children shown in the red square on the PCA plot. It also allows quantitative assessment of similarity/dissimilarity using the cluster separation metric.

A, Expanded view of the gray shaded area in PCA plot of Figure 4, with similar pair (black line) and similar group (red dash line) of children highlighted. B, dendrogram plot showing cluster relationships, cluster 17 corresponds to the red-dash highlighted area of the PCA plot. C and D, Acceleration magnitude plots for the pair of children highlighted by the black outline in the PCA plot.

Longitudinal study of the PCA plot provides insight on the daily activity patterns of the children. As an example, two children with differing patterns of play are shown highlighted in Figure 6. Child A exhibits highly varied activity profiles with large day-to-day variance in the level and the time pattern of physical motion, whereas child B displays a tight cluster of points from consistent daily activity patterns. Analysis of the varying activity profiles across a week also point to social influences on play. The extremely consistent play pattern of child B is disrupted on the Friday of the study week, and they have a much reduced activity level on this day. Inspection of the PCA plot shows that the upper left region, in which the Friday play data of child B sits, is dominated by a cluster of other data points from Friday activity traces for children from the same class. The time-dependent acceleration traces for all of this group show an extended period of inactivity between the 15- and 25-min points of the play session (see additional figures, Supplemental Digital Content 1, Appendix—supplementary information, Thus, there is strong circumstantial evidence that the altered play pattern of child B is due to the influence of their peer group.

Motion-metrics PCA with daily activity of two children highlighted (child A—red, child B—green). Black-filled circles represent Friday activity points of six children in same year (year 6) as child B.


The aim of this study was to demonstrate automated, quantitative assessment of children’s movement during play. There is a growing appreciation of the importance of the quality of activity in developing movement competence (9) and automated assessment of various movement tasks, based on signal feature extraction from wearable sensors, has been reported (36,37). Recognition and classification of activity type have also been achieved using Machine Learning algorithms (38,39). Although these approaches provide enhanced metrics on activity, over and above simple quantification, they are based on a premise that there exists a stable and recognizable movement pattern associated with each activity category, for example, walk, run, skip, and so on. For children at play, it is debatable whether such pattern standards exist. They exhibit an almost unlimited range of movement and even in core motions, such as walking will display highly varied patterns, both at an individual level in step-to-step variance and at population level in the changing walk style across the cohort. In recognition of this, we have developed an alternative method for activity profiling, based on identifying when movement takes place rather than when acceleration is produced. This follows other quantitative techniques in being focused on discrete motion events rather than continuous acceleration-based metrics. However, it offers a novel alternative when characterizing these events as it implements indirect measurement by comparison to a reference standard, rather than direct extraction of data from the child’s motion signal. The advantage here is that because the acceleration patterns displayed by the children can be of any form, movement is no longer constrained to fit to a preordained pattern. Benchmarked quantification is maintained as the extraction of metrics is always in reference to the known motif, which becomes the yardstick for interpretation.

As all metrics stem from comparison to a single motif this approach provides robust data that support comparison across a cohort and across different study days. By resolving movement into motion events of short duration, the technique also provides multiparameter descriptions of each child’s play session, and this allows multivariate profiling of the cohort. All of the moment-to-moment detail of the varied play activity is captured and can be mapped, using a dimensional reduction algorithm, into a map of all activity profiles. This allows visual inspection of the variation or uniformity in activity level and identification of clusters of like individuals who display similar activity patterns. This clustering may also point to social influences upon play, as closely mapped children have very similar acceleration time traces, and this suggests common play during the measured period. The activity map also shows temporal change across a longitudinal study and could be a powerful analysis tool in the case of intervention, where coordinate position on the map shows comparative performance between individuals, preintervention and postintervention.

In the work presented, the reference standard chosen was a walk step, but other motifs could be used to give activity profiling in relation to a running stride, hop step, arm movement, or similar. It is important to note that the motif determines the values of the extracted activity parameters but does not change the form of the measured motion signal. Thus, if an alternative motif pattern is used, the values of Δd and Δt for each event will be different but the density of movement events and the comparative relationships between children will be unaltered. In this respect, the motif acts as a filter through which we view the children’s movement; changing it provides a different perspective of the same underlying activity topography.

J. N. R. undertook this work while on a research visit funded by the Erasmus+ Credit Mobility Programme (2017-1-PT01-KA103–035245), C.M.B. was funded under a UK EPSRC Doctoral Training Grant. The authors thank the children of the city and county of Swansea who participated in the study. The results of the present study do not constitute endorsement by ACSM. The results are presented clearly, honestly, and without fabrication, falsification, or inappropriate data manipulation.

Present address for C. T. C., Centre for Sport, Exercise and Life Sciences, Coventry University, Priory Street, Coventry, CV1 5FB, United Kingdom.


1. Kohl HW, Craig CL, Lambert EV, et al. The pandemic of physical inactivity: global action for public health. Lancet. 2012;380(9838):294–305.
2. Blair SN. Physical inactivity: the biggest public health problem of the 21st century. Br J Sports Med. 2009;43(1):1–2.
3. Landry BW, Driscoll SW. Physical activity in children and adolescents. PM&R. 2012;4(11):826–32.
4. Janssen I, LeBlanc AG. Systematic review of the health benefits of physical activity and fitness in school-aged children and youth. Int J Behav Nutr Phys Act. 2010;7:40.
5. Sallis JF, Prochaska JJ, Taylor WC. A review of correlates of physical activity of children and adolescents. Med Sci Sports Exerc. 2000;32(5):963–75.
6. Willenberg LJ, Ashbolt R, Holland D, et al. Increasing school playground physical activity: a mixed methods study combining environmental measures and children’s perspectives. J Sci Med Sport. 2010;13(2):210–6.
7. Ridgers ND, Stratton G, Fairclough SJ, Twisk JW. Long-term effects of a playground markings and physical structures on children’s recess physical activity levels. Prev Med. 2007;44(5):393–7.
8. Lai SK, Costigan SA, Morgan PJ, et al. Do school-based interventions focusing on physical activity, fitness, or fundamental movement skill competency produce a sustained impact in these outcomes in children and adolescents? A systematic review of follow-up studies. Sports Med. 2014;44(1):67–79.
9. Myer GD, Faigenbaum AD, Edwards NM, Clark JF, Best TM, Sallis RE. Sixty minutes of what? A developing brain perspective for activating children with an integrative exercise approach. Br J Sports Med. 2015;49:1510–6.
10. Ridgers ND, Salmon J, Parrish A-M, Stanley RM, Okely AD. Physical activity during school recess. Am J Prev Med. 2012;43(3):320–8.
11. Sirard JR, Pate RR. Physical activity assessment in children and adolescents. Sports Med. 2001;31(6):439–54.
12. Ott AE, Pate RR, Trost SG, Ward DS, Saunders R. The use of uniaxial and triaxial accelerometers to measure children’s “free-play” physical activity. Pediatr Exerc Sci. 2000;12(4):360–70.
13. Ridgers ND, Stratton G, Fairclough SJ. Physical activity levels of children during school playtime. Sports Med. 2006;36(4):359–71.
14. Brandes M, Zijlstra W, Heikens S, van Lummel R, Rosenbaum D. Accelerometry based assessment of gait parameters in children. Gait Posture. 2006;24(4):482–6.
15. Yang CC, Hsu YL. A review of accelerometry-based wearable motion detectors for physical activity monitoring. Sensors. 2010;10(8):7772–88.
16. Welk GJ, Corbin CB, Dale D. Measurement issues in the assessment of physical activity in children. Res Q Exerc Sport. 2000;71(2 Suppl):59–73.
17. Kozey SL, Lyden K, Howe CA, Staudenmayer JW, Freedson PS. Accelerometer output and MET values of common physical activities. Med Sci Sports Exerc. 2010;42(9):1776–84.
18. Gavrila DM. The visual analysis of human movement: a survey. Comput Vis image Underst. 1999;73(1):82–98.
19. Bobick AF, Davis JW. The recognition of human movement using temporal templates. IEEE Trans Pattern Anal Mach Intell. 2001;23(3):257–67.
20. Zhou F, De la Torre F. Generalized canonical time warping. IEEE Trans Pattern Anal Mach Intell. 2016;38(2):279–94.
21. Halilaj E, Rajagopal A, Fiterau M, Hicks JL, Hastie TJ, Delp SL. Machine learning in human movement biomechanics: best practices, common pitfalls, and new opportunities. J Biomech. 2018;81:1–11.
22. De Vries SI, Garre FG, Engbers LH, Hildebrandt VH, Van Buuren S. Evaluation of neural networks to identify types of activity using accelerometers. Med Sci Sports Exerc. 2011;43(1):101–7.
23. Wang J, Chen Y, Hao S, Peng X, Hu L. Deep learning for sensor-based activity recognition: a survey. Pattern Recognit Lett. 2019;119:3–11.
24. Bonomi AG, Goris AH, Yin B, Westerterp KR. Detection of type, duration, and intensity of physical activity using an accelerometer. Med Sci Sports Exerc. 2009;41(9):1770–7.
25. Willetts M, Hollowell S, Aslett L, Holmes C, Doherty A. Statistical machine learning of sleep and physical activity phenotypes from sensor data in 96,220 UK biobank participants. Sci Rep. 2018;8(1):7961.
26. van Kuppevelt D, Heywood J, Hamer M, Sabia S, Fitzsimons E, van Hees V. Segmenting accelerometer data from daily life with unsupervised machine learning. Buchowski MS, editor. PLoS One. 2019;14(1):e0208692.
27. Riddoch CJ, Mattocks C, Deere K, et al. Objective measurement of levels and patterns of physical activity. Arch Dis Child. 2007;92(11):963–9.
28. Bringolf-Isler B, Grize L, Mäder U, Ruch N, Sennhauser FH, Braun-Fahrländer C. Assessment of intensity, prevalence and duration of everyday activities in Swiss school children: a cross-sectional analysis of accelerometer and diary data. Int J Behav Nutr Phys Act. 2009;6(1):50.
29. Mota J, Silva P, Santos MP, Ribeiro JC, Oliveira J, Duarte JA. Physical activity and school recess time: differences between the sexes and the relationship between children’s playground physical activity and habitual physical activity. J Sports Sci. 2005;23(3):269–75.
30. Barnes CM, Clark CC, Holton MD, Stratton G, Summers HD. Quantitative time-profiling of children’s activity and motion. Med Sci Sports Exerc. 2017;49(1):183–90.
31. Kim Y, Beets MW, Welk GJ. Everything you wanted to know about selecting the “ right” Actigraph accelerometer cut-points for youth, but...: a systematic review. J Sci Med Sport. 2012;15(4):311–21.
32. Barnes CM, Clark CC, Rees P, Stratton G, Summers HD. Objective profiling of varied human motion based on normative assessment of magnetometer time series data. Physiol Meas. 2018;39(4):045007.
33. Berndt DJ, Clifford J. Using dynamic time warping to find patterns in time series. In: KDD Workshop. Seattle, WA; 1994. pp. 359–70.
34. Müller M. Dynamic time warping. Inf Retr Music motion. 2007;69–84.
35. Ainsworth BE, Haskell WL, Leon AS, et al. Compendium of physical activities: classification of energy costs of human physical activities. Med Sci Sports Exerc. 1993;25(1):71–80.
36. Bisi MC, Pacini Panebianco G, Polman R, Stagni R. Objective assessment of movement competence in children using wearable sensors: an instrumented version of the TGMD-2 locomotor subtest. Gait Posture. 2017;56:42–8.
37. Masci I, Vannozzi G, Bergamini E, Pesce C, Getchell N, Cappozzo A. Assessing locomotor skills development in childhood using wearable inertial sensor devices: the running paradigm. Gait Posture. 2013;37(4):570–4.
38. Fergus P, Hussain A, Hearty J, et al. A machine learning approach to measure and monitor physical activity in children to help fight overweight and obesity. Lect Notes Comput Sci. 2015;9226:676–88.
39. De Vries SI, Engels M, Garre FG. Identification of children’s activity type with accelerometer-based neural networks. Med Sci Sports Exerc. 2011;43(10):1994–9.


Supplemental Digital Content

© 2020 American College of Sports Medicine