Share this article on:

Predicting Human Movement with Multiple Accelerometers Using Movelets


Medicine & Science in Sports & Exercise: September 2014 - Volume 46 - Issue 9 - p 1859–1866
doi: 10.1249/MSS.0000000000000285
SPECIAL COMMUNICATIONS: Methodological Advances

Purpose The study aims were 1) to develop transparent algorithms that use short segments of training data for predicting activity types and 2) to compare the prediction performance of the proposed algorithms using single accelerometers and multiple accelerometers.

Methods Sixteen participants (age, 80.6 yr (4.8 yr); body mass index, 26.1 kg·m−2 (2.5 kg·m−2)) performed 15 lifestyle activities in the laboratory, each wearing three accelerometers at the right hip and left and right wrists. Triaxial accelerometry data were collected at 80 Hz using ActiGraph GT3X+. Prediction algorithms were developed, which, instead of extracting features, build activity-specific dictionaries composed of short signal segments called movelets. Three alternative approaches were proposed to integrate the information from the multiple accelerometers.

Results With at most several seconds of training data per activity, the prediction accuracy at the second-level temporal resolution was very high for lying, standing, normal/fast walking, and standing up from a chair (the median prediction accuracy ranged from 88.2% to 99.9% on the basis of the single-accelerometer movelet approach). For these activities, wrist-worn accelerometers performed almost as well as hip-worn accelerometers (the median difference in accuracy between wrist and hip ranged from −2.7% to 5.8%). Modest improvements in prediction accuracy were achieved by integrating information from multiple accelerometers.

Discussion and Conclusions It is possible to achieve high prediction accuracy at the second-level temporal resolution with very limited training data. To increase prediction accuracy from the simultaneous use of multiple accelerometers, a careful selection of integrative approaches is required.

1Department of Biostatistics, The Johns Hopkins University, Baltimore, MD; 2Department of Social Medicine, University of Maastricht, Maastricht, THE NETHERLANDS; 3Institute of Sports Science and Clinical Biomechanics, University of Southern Denmark, Odense, DENMARK; 4Department of Epidemiology, University of Pittsburgh, Pittsburgh, PA; and 5Laboratory of Epidemiology, Demography, and Biometry, National Institute on Aging, Bethesda, MD

Address for correspondence: Bing He, ScM, The Johns Hopkins University, E3148, 615 N. Wolfe St., Baltimore, MD 21205-2179; Email:

Submitted for publication August 2013.

Accepted for publication January 2014.

The objective and detailed characterization of daily physical activity is crucial for research studies where physical activity is an exposure or outcome. Researchers have increasingly relied on accelerometers for measuring physical activity in observational studies and clinical trials (1,5,6,9,18). A fundamental question is how to decipher and interpret the accelerometry signals into meaningful information such as activity intensity, energy expenditure, and movement types (20–22,25).

Various methods have been previously suggested for recognition of activity types, including linear/quadratic discriminant analysis (13), hidden Markov chains (10), artificial neural networks (19,23,24), support vector machines (12,15) and combined methods (17,28). Bao and Intille (3) and Preece et al. (14) reviewed and evaluated methods used in the classification of normal activities. The major limitations of the other methods are as follows: 1) they usually require at least a 1-min window for a feature extraction approach; 2) they do not capture finer movements that last less than 1 min, such as falling or standing up from a chair; and 3) the prediction process is usually hard to understand and interpret. For example, when the prediction algorithm predicts correctly, it remains unclear which features contribute and what their relative contributions to activity discrimination is; similarly, when the prediction algorithm fails, it is unclear exactly what features created the activity prediction confusion.

This article provides movelet-based prediction algorithms that can address these limitations. The methodology is fully transparent, easy to understand, and requires minimal training data. The intuition behind is that movements with similar accelerometry patterns are likely to correspond to the same type of activity. The single-accelerometer movelet approach proposed by Bai et al. (2) was developed for one triaxial accelerometer, and this article extends it to multiple accelerometers. A movelet is the collection of the three acceleration time series recorded by the device in a window of given length, say 1 s. The sets of movelets constructed from the accelerometry data with annotated labels are organized by activity types, i.e., “chapters,” which play the role of accelerometry “dictionaries” for different activity types. Predictions of accelerometry data without annotated labels are obtained by identifying the chapter that is most similar to the data in terms of distance; the similarity is quantified by the standard Euclidean distance between vectors. This can be extended to multiple accelerometers in at least two ways: 1) by building separate movelet dictionaries and then combining predictions using voting or sequential decisions, or 2) by designing a joint dictionary, where a movelet is a collection of nine time series (three for each accelerometer). The movelets idea (2) uses a dictionary learning approach that has been used extensively in speech recognition and is likely to be one of the fundamental algorithms for movement recognition in game consoles. The “shapelets” algorithms also use dictionary learning to recognize shapes of objects from images and videos (7,27). A recent development of the “fast shapelets” method includes an application to accelerometry-based activity type recognition (16). A recently published conference article (8) proposed a movement prediction framework based on accelerometry data similar to the one in Ref. (2). Both papers (8,16) were published a year later than Ref. (2). Zheng et al. (29) divided accelerometry signals into nonoverlapping windows and adopted a multiscale ensemble method to relax the window size specification; our movelets-based methods decompose accelerometry signals into overlapping windows (i.e., movelets) and thus accommodate that different activities occur at different temporal scales. Here we focus on interpretability and generalization of the movelets-based approach in Ref. (2) from one to multiple accelerometers, on studying what types of integration approaches work best, on better characterizing complex movements, and on applying the methods to some of the most complex in-lab data available.

Back to Top | Article Outline



Sixteen older adults (age, 80.6 yr (4.8 yr); body mass index, 26.1 kg·m−2 (2.5 kg·m−2); 50.0% female) enrolled in the Study of Energy and Aging Pilot were invited to participate in an additional study visit to validate hip and wrist accelerometry. Of the 38 participants in the Study of Energy and Aging, 16 were interested and completed this ancillary study. This study was approved by the Institutional Review Board of the University of Pittsburgh and National Institute on Aging and all participants gave written informed consent.

Back to Top | Article Outline

Data collection

The participants were instructed to perform 15 different types of activities according to a protocol in a research clinic, including lying still face-up (alias: lying), standing still (alias: standing), washing plates (alias: washDish), kneading a ball of dough (alias: dough), putting jacket on (alias: dressing), folding towels and stacking them nearby (alias: foldTowel), vacuuming carpet (alias: vacuum), simulated shopping (alias: shop), writing (alias: write), dealing cards (alias: cards), standing up from a chair and sitting back down (alias: chairStand), walking at a normal speed with (alias: normalWalk Swing) and without arm swing (alias: normalWalk noSwing), and walking at a fast speed with (alias: fastWalk Swing) and without arm swing (alias: fastWalk noSwing). In the rest of the article, the activity types are referred to by their aliases. Each participant wore three ActiGraph GT3X+ devices simultaneously, which were located at the right hip, right wrist, and left wrist. For each accelerometer, the collected data contain a triaxial time series of accelerations expressed in units of gravity, i.e., g = 9.81 m·s−2. The three axes are labeled as “SuperiorInferior,” “AnteriorPosterior,” and “LeftRight” according to the axes’ directions with respect to the participants in a standard standing position. The data were collected at a sampling frequency of 80 Hz. On the basis of the protocol and the start/end time for each activity, a time series of activity type labels is constructed to annotate the accelerometry data.

Back to Top | Article Outline

ActiGraph GT3X+

The ActiGraph GT3X+ device is a triaxial acceleration sensor developed by ActiGraph, LLC. It is a triaxial ±6g seismic acceleration sensor housed in a small (4.6 cm × 3.3 cm × 1.5 cm) light weight (19 g) encasing. The user-defined sampling rate of the ActiGraph GT3X+ can range from 30 to 100 Hz in 10-Hz increments.

Back to Top | Article Outline

Single-accelerometer movelet approach

We first review the movelets approach (2) developed for predicting activity type using a single accelerometer. The basic idea is to decompose the triaxial time series into movelets, overlapping short segments of data representing a given activity. The prediction of the activity type is based upon the similarity (distance) between the movelets that are not labeled and the ones that are. The entire process is similar to having a dictionary of words (movelets) with their associated meaning (labels). Given a new word (unlabeled movelet), the procedure simply requires looking up the word with the closest meaning in the dictionary (labeled movelets) and assigning the corresponding label to the new word (unlabeled movelet). The basic idea is simple: (a) cut the time series that correspond to known and unknown activities into small overlapping 1-s intervals; and (b) match each time series corresponding to an unknown activity to the most similar time series corresponding to a known activity. For notation and complete description of the single-accelerometer movelet approach, see Ref. (2).

Back to Top | Article Outline

Integrative movelets approaches

It is reasonable to expect that accelerometers worn at different body locations contain nontrivial complementary information about movements. However, it is not clear whether and how this information can be combined to augment the prediction performance. Here we propose three practical approaches for combining information. Combining classifiers to improve prediction is an intensely studied topic in statistics and machine learning. In statistics, Breiman (4) and Wolpert (26) discussed model stacking and averaging. Lam (11) provided a review of different methods of combining classifiers in machine learning. We propose three easy to understand and scale up integrative approaches: movelets voting, movelets decision tree, and expanded movelets. The movelets voting approach integrates the three accelerometers by allowing the single-accelerometer movelet approach to vote for the activity types and then by accepting the majority votes. The movelets decision tree builds up a simple hierarchy of decisions based on movelets. The hip-worn accelerometer first discriminates among the top-level groups of activities, followed by the low-level prediction using wrist-worn accelerometers for specific activity types. The expanded movelets simply redefines movelets as a combination of nine 1-s time series (3 for each accelerometer); weights can be imposed to accelerometers attached to different positions. In this study, equal weights for hip and right and left wrist accelerometers were adopted.

Because of definition ambiguity, the breaks between two successive activities as well as the transition periods at the beginning and end of each activity are removed from the data. In the following analysis, data for a half-minute period for each activity of each participant are used as the testing data. The length of the movelet H is taken to be 80 observations, i.e., 1-s segment. A 1-s window is appropriate because it contains enough information to identify a movement without including much redundant information. A dictionary of 15 chapters corresponding to each of the 15 activities in the study is created for each accelerometer and each participant. For activities with an explicit start and end, such as chairStand, one replicate is used as training data. For movements with periodic features, such as normalWalk and fastWalk, a 2-s segment is utilized as training data. For other continuous movements without explicit periodicity, a segment with a length 2.5 s is used as training data. To make comparison, we kept the training and testing data consistent across different methods for each subject. Movelet dictionaries are subject specific in the single movelet approach. In the integrative movelets approaches, movelet dictionaries are subject-accelerometer specific; for each accelerometer and each subject, we have constructed a dictionary of activities on the basis of the training data. Dictionaries are distinct for each subject to account for the individual variations in movement patterns across subjects. For one subject, dictionaries were designed to be accelerometer specific to capture different characteristics of movements recorded from different replacements. In the future, we will explore the possibility of using the dictionary of one or multiple subjects to predict the movement of a different subject.

Back to Top | Article Outline


Figure 1 displays the boxplot of the prediction accuracy, i.e., the proportions of correctly classified observations, for 15 activity types of 16 participants using hip-worn, right wrist-worn, and left wrist-worn accelerometers. The X axis indicates the various activities, and the Y axis corresponds to the proportion of correctly classified observations. Different colors were used for each prediction approach: 1) red for movelets based on hip accelerometers; 2) blue for movelets based on right wrist accelerometers; 3) green for movelets based on left wrist accelerometers; 4) yellow for the expanded movelets approach; 5) orange for the movelets voting approach; and 6) grey for the movelets decision tree approach. Note that a new activity type, normalWalk combined, merges normalWalk Swing and normalWalk noSwing into one category; and fastWalk combined is fastWalk Swing plus fastWalk noSwing. The activity labels on the X axis are ordered with respect to median prediction accuracy of the hip-worn accelerometer (red). Table 1 presents the median and interquartile range of prediction accuracies across participants for different activity types.





Back to Top | Article Outline

Results for single-accelerometer movelet approach

For resting activities (lying and standing), all accelerometers provide accurate predictions. For normalWalk combined and fastWalk combined, the wrist-worn accelerometers perform as well as the hip-worn accelerometers. The right wrist-worn and left wrist-worn accelerometers outperform hip-worn accelerometers in predicting normalWalk Swing, normalWalk noSwing, fastWalk Swing, fastWalk noSwing, and chairStand. Write and cards belong to the group of upper body activities while sitting. Accelerometers worn at three different positions yield very accurate predictions for writing (see the red, blue, and green boxplots in Figure 1 corresponding to write). Right wrist accelerometers falsely predict on average 10.2% of cards to be write. For cards, left wrist-worn accelerometers provide higher median prediction accuracy than right wrist-worn accelerometers, and both outperform the hip-worn accelerometers. Hip-worn accelerometers do not record the subtle movements of the hands and often incorrectly classified cards as dough (5.4% average across participants) and foldTowel (5.3% average across participants). On average, 12.8%, 13.1%, and 9.4% of dressing is falsely classified as foldTowel based on the hip-worn and right- and left wrist-worn accelerometers, respectively. For activities dough, washDish, vacuum, dressing, foldTowel, and shop, all three accelerometers show lower median prediction accuracy and larger variability across participants.

Back to Top | Article Outline

Results for integrative movelets approaches

For the activities lying, normalWalk combined, write, stand, normalWalk Swing, fastWalk Swing, and cards, the expanded movelets approach yields the highest prediction accuracy with the least variability across participants. A substantial increase in prediction accuracy for cards is observed for the expanded movelets approach. To provide a visual representation, Figure 2 shows the matching processes of two participants for whom the expanded movelets approach provides better prediction for cards than all the single-accelerometer movelet approach on its own. The movelets voting approach is inferior to either one of the single-accelerometer movelet approach or the expanded movelets approaches for all the activities with the exception of dough, washDish, vacuum, and shop. The movelets decision tree provides relatively good performance for lying, write, and stand. For other activities, it tends to underperform.



Back to Top | Article Outline


We conducted a four-fold cross-validation to further evaluate the proposed methods. More precisely, every time, we used another part of the data for training and testing and have investigated the sensitivity of the proposed approach. For activities lying, walking, stand, chairStand, and write, the within-subject variability of the prediction accuracy is much smaller than the between-subject variability. For most of the household activities, the within-subject variability of prediction accuracy is larger and the average prediction accuracy over folds is also generally low; this is due to the ambiguity of submovements that are shared among multiple household activities. Results suggest that for activities lying, walking, stand, chairStand, and write, the movelet-based approaches provide highly consistent results when using different training sets.

Back to Top | Article Outline


Discussion on movelets approaches

This work provides movelet-based prediction of activity type at the second-level temporal resolution using triaxial accelerometers placed at the right hip and left and right wrists. Compared with feature extraction-based methods, the movelet-based methods have two advantages. First, instead of extracting features, the movelet algorithm preserves the original signal that can be visually inspected. The matching process between known and unknown movelets mimics the natural human pattern recognition, which makes the process transparent. If prediction fails in a 1-s interval, human visual inspection can usually reveal the reason. Second, movelets can achieve high prediction accuracy at the second level even using very limited training data (several seconds per activity type). This is the first approach that can achieve such performance although remaining fully transparent and easy to understand.

To illustrate how these advantages work in practice, we present several examples here. First, it was found that wrist-worn accelerometers predict walking better than hip-worn accelerometers. This is unexpected and counterintuitive, because these movements are fundamentally performed by producing lower body acceleration. As a second example, the results showed that the accelerometers worn at the left wrist outperformed the one worn at the right wrist in recognizing cards, which is also unexpected. To investigate these, Figure 3 displays the matching processes for normalWalk, fastWalk, and cards. Activity labels are coded in different colors; the annotated labels and predicted labels are plotted in parallel accompanying the original signals. Given a time point, if the annotated label and the predicted label are of the same color, then the activity type is correctly predicted.



Consider the matching process for walking with and without arm swing as measured at the hip (first row, left panel) and right wrist (second row, left panel) in Figure 3. As the person transitions from normal walk without arm swing to normal walk with arm swing, the time series associated with hip movement do not display visually observable changes. In contrast, the wrist accelerometry indicates a strong change. Most interestingly, the blue accelerometry curve shifts to a much higher level than before, probably because of the change in the angle of the accelerometer. Such strong angle changes can be easily observed and detected using movelets and should explain how information is being combined. The single-accelerometer movelet approach is confused between the two types of movements when using the hip data only. In contrast, the predicted labels based on wrist data are almost perfectly accurate. A similar story holds for fast walking with or without arm swing (row 3 and 4 panels). Now examine the right panels in the first and second rows of Figure 3. It is shown that the right wrist accelerometer yields more variable signal than the left wrist one, which makes it harder to distinguish the right wrist accelerometer signals from other upper body activities. As a third example, the prediction accuracy for the upper limb activities while standing was lower and yielded larger variability across participants. This probably happens because of the high level of overlap in movement and ambiguity of some submovements across labeled activities. At the same time, the raw signals of these activities are also similar with visual inspection of the matching process.

Three integrative movelets approaches were proposed to incorporate information from multiple accelerometers. The expanded movelets approach yields the best overall performance among the three integrative movelets approaches. Movelets voting is conceptually straightforward. Because the integration of information occurs after using single-accelerometer movelet approach for each accelerometer, it can simultaneously process acceleration information from different sources and save computing time. At the same time, integration with majority votes of categorical activities labels loses some of the rich information embedded in the original signals. The expanded movelets approach merges all available information and provides the flexibility of weighting different information sources. This method yields exceptional prediction in well-controlled environments, although it may be more prone to errors when one of the devices malfunctions or moves to a different position on the body.

The weights on the sources can be assigned depending on activity types. Determining optimal weights for different sources is not covered in this article, but it will be investigated in the future. The movelets decision tree approach integrates information adaptively. An important assumption is that the top level of the decision tree is well designed to provide coarse discrimination between activity groups, whereas lower level decisions are well designed to make within-group predictions. In the movelets decision tree approach, the prediction error in the first level is propagated into the second level. Thus, designing the tree hierarchy is rather delicate and, likely, application specific. To increase prediction accuracy from the simultaneous use of multiple accelerometers, a careful selection of integrative algorithms is required.

Back to Top | Article Outline

Methodological implications

The order of magnitudes of local average accelerations among the three axes is crucial for detecting and differentiating various postures; this happens because the order is a proxy of the orientation of the device relative to earth gravity. For example, our results showed that high accuracy was achieved for lying and standing. The visual inspection of the matching process reveals that the main difference between the accelerometry signals for lying and standing is that the local average of individual time series is different in magnitude and rank. While standing still, the gravity would appear as acceleration mainly along the AnteriorPosterior axis. Gravity affects differently each axis, and the size of the effect depends fundamentally on the angles the axes of the accelerometer form with the gravity direction. Although far from being a perfect proxy for the position, this is enough to differentiate between standing and lying. This is a case where the variability of time series along their long-term averages is of secondary importance, whereas the discrimination between the two resting positions is done by the shift in the relative magnitude of the mean functions.

Our findings have potential implication on the accelerometers’ placement decisions in epidemiological studies. First, both accelerometers worn at the dominant hand and nondominant hand can predict lying, standing, normal walking, and fast walking as well as the hip-worn accelerometers. This provides support for using wrist-worn accelerometers. Second, handedness should also be a consideration in accelerometer placement, which can affect prediction performance. Our results showed that in activities like cards where the dominant hand moves more, the nondominant hand accelerometers yielded higher prediction accuracy. Third, the decision of placement should be customized to the study goals. The results showed that hip-worn accelerometers performed poorly in distinguishing walking with arm swing versus walking without arm swing, whereas wrist-worn accelerometers yielded good results. Consider a scenario when investigators decide to use hip-worn accelerometers and are interested in distinguishing between different types of walking. It will be quite difficult to differentiate between walking normally and walking carrying a small object (no arm swing). Thus, it seems reasonable to simply define a label called normal walking that includes both arm swing and no swing. Alternatively, a wrist accelerometer could be used instead or in addition to the hip one. It was also shown that trunk movements such as walking are easily recognized by all accelerometers, whereas finer movements that occur at the extremities of the human body are predominantly captured by wrist-worn accelerometers.

Back to Top | Article Outline


We have seen a trend that more and more devices are now designed to output raw triaxial data and more and more researchers have started to collect the raw data. There is also increasing interest in evaluating data that are being lost by analyzing the coarse activity data at the hour or day level versus the raw accelerometry data recorded at the subsecond level. The movelets approaches proposed here provide a self-consistent and transparent framework for thinking about and quantifying the data at the subsecond level. The movelets approaches allow the scientist to understand the complex measurement, have access to the entire processing pipeline, and access different levels of data compression via reproducible code and verifiable results.

In this study, all movement predictions were obtained from participants in the research clinic wearing devices installed by trained technicians. The activities are intended to represent activities that happen in the participant’s own environment. As much as one tries to standardize laboratory experiments, the data are likely to provide only a partial snapshot of the heterogeneous activities individuals perform in their own houses. It remains unclear how in-lab data prediction algorithms perform in free-living environments, especially in the absence of labeled in-home data on hundreds of individuals. Also, we have not yet investigated how well methods could be trained on one or multiple participants and then applied to other participants. However, these issues provide opportunities for future work.

This research was funded in part by the Intramural Research Program of the National Institute on Aging. Dr. Crainiceanu was supported by the National Institute of Neurological Disorders and Stroke under Award Numbers R01NS060910 and R01NS085211 and by the National Institute of Mental Health of the National Institutes of Health under Award Number R01MH095836.

This work was also supported by the National Institute of Health through the funded Study of Energy and Aging Pilot (RC2AG036594), Pittsburgh Claude D. Pepper Older American’s Independence Center Research Registry (NIH P30 AG024826), a National Institute on Aging Professional Services Contract (HHSN271201100605P), a University of Pittsburgh Department of Epidemiology Small Grant, and the National Institute on Aging Training Grant (T32AG000181).

This work represents the opinions of the researchers and not necessarily that of the granting organizations.

All the authors have no conflict of interest.

The present study does not constitute endorsement by the American College of Sports Medicine.

Back to Top | Article Outline


1. Atienza AA, King AC. Comparing self-reported versus objectively measured physical activity behavior: a preliminary investigation of older Filipino American women. Res Q Exerc Sport. 2005; 76 (3): 358–62.
2. Bai J, Goldsmith J, Caffo BS, Glass T, Crainiceanu CM. Movelets: a dictionary of movement. Electron J Stat. 2012; 6: 559–78.
3. 3. Bao L, Intille SS. Activity recognition from user-annotated acceleration data. In: Proceedings of the 2nd International Conference on Pervasive Computing, 2004; 1–17.
4. Breiman L. Bagging predictors. Mach Learn. 1996; 24 (2): 123–40.
5. Bussmann JB, Martens WL, Tulen JH, Schasfoort FC, van den Berg-Emons HJ, Stam HJ. Measuring daily behavior using ambulatory accelerometry: the activity monitor. Behav Res Methods Instrum Comput. 2001; 33 (3): 349–56.
6. Choi L, Liu Z, Matthews CE, Buchowski MS. Validation of accelerometer wear and nonwear time classification algorithm. Med Sci Sports Exerc. 2011; 43 (2): 357–64.
7. 7. Hills J, Lines J, Baranauskas E, Mapp J, Bagnall A. Time series classification with shapelets. J Data Mining Knowledge Disc. 2014; 28 (4): 851–81.
8. Hu B, Chen Y, Zakaria J, Ulanova L, Keogh E. Classification of Multi-Dimensional Streaming Time Series by Weighting each Classifiers Track Record. Dallas (TX): International Conference on Data Mining (ICDM); 2013.
9. Kozey-Keadle S, Libertine A, Lyden K, Staudenmayer J, Freedson P. Validation of wearable monitors for assessing sedentary behavior. Med Sci Sports Exerc. 2011; 43 (8): 1561.
10. Krause A, Sieiorek DP, Smailagic A, Farringdon J. Unsupervised, dynamic identification of physiological and activity context in wearable computing. In: Proceedings of the 7th International Symposium on Wearable Computers. White Plains (NY): IEEE Computer Society; 2003. pp. 88–97.
11. Lam L. Classifier combinations: implementations and theoretical issues. Lect Notes Comput Sci. 2000; 1847: 77–86.
12. Mannini A, Intille SS, Rosenberger M, Sabatini AM, Haskell W. Activity recognition using a single accelerometer placed at the wrist or ankle. Med Sci Sports Exerc. 2013; 45 (11): 2193–203.
13. Pober DM, Staudenmayer J, Raphael C, Freedson PS. Development of novel techniques to classify physical activity mode using accelerometers. Med Sci Sports Exerc. 2006; 38 (9): 1626–34.
14. Preece S, Goulermas J, Kenney L, Howard D, Meijer K, Crompton R. Activity identification using body-mounted sensors: a review of classification techniques. Physiol Meas. 2009; 30 (4): R1–33.
15. Qian H, Mao Y, Xiang W, Wang Z. Recognition of human activities using SVM multi-class classifier. Pattern Recognit Lett. 2010; 31 (2): 100–11.
16. Rakthanmanon T, Keogh E. Fast shapelets: a scalable algorithm for discovering time series shapelets. In: Proceedings of the 13th SIAM International Conference on Data Mining (SDM), 2013, Austin (TX).
17. Ravi N, Dekar N, Mysore P, Littman ML. Activity recognition from accelerometer data. In: Proceedings of the Seventeenth Conference on Innovative Applications of Artificial Intelligence. 2005; 3: 1541–6.
18. Sirard JR, Trost SG, Pfeiffer KA, Dowda M, Pate RR. Calibration and evaluation of an objective measure of physical activity in pre-school children. J Phys Act Health. 2005; 2 (3): 345–7.
19. Staudenmayer J, Pober D, Crouter S, Bassett D, Freedson P. An artificial neural network to estimate physical activity energy expenditure and identify physical activity type from an accelerometer. J Appl Physiol. 2009; 107 (4): 1300–7.
20. Staudenmayer J, Zhu W, Catellier DJ. Statistical considerations in the analysis of accelerometry-based activity monitor data. Med Sci Sports Exerc. 2012; 44 (1 Suppl): S61–7.
21. Troiano RP, Berrigan D, Dodd KW, Msse LC, Tilert T, McDowell M. Physical activity in the united states measured by accelerometer. Med Sci Sports Exerc. 2008; 40 (1): 181–8.
22. Trost SG, McIver KL, Pate RR. Conducting accelerometer-based activity assessments in field-based research. Med Sci Sports Exerc. 2005; 37 (11 Suppl): S531.
23. Trost SG, Wong WK, Pfeier KA, Zheng Y. Artificial neural networks to predict activity type and energy expenditure in youth. Med Sci Sports Exerc. 2012; 44: 1801–9.
24. Vries De SI, Garre FG, Engbers LH, Hildebrt VH, Buuren SV. Evaluation of neural networks to identify types of activity using accelerometers. Med Sci Sports Exerc. 2011; 43: 101–7.
25. Welk GJ, Blair SN, Jones S, Thompson RW. A comparative evaluation of three accelerometry-based physical activity monitors. Med Sci Sports Exerc. 2000; 32: 489–97.
26. Wolpert DH. Stacked generalization. Neural Netw. 1992; 5: 241–59.
27. Ye L, Keogh E. Time series shapelets: a new primitive for data mining. In: Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2009, Paris France.
28. Zhang S, Rowlands AV, Murray P, Hurst TL. Physical activity classification using the GENEA wrist-worn accelerometer. Med Sci Sports Exerc. 2012; 44 (4): 742–8.
29. Zheng Y, Wong WK, Guan X, Trost S. Physical activity recognition from accelerometer data using a multi-scale ensemble method. In: Proceedings of the Twenty-Fifth Innovative Applications of Artificial Intelligence Conference. 2013.


© 2014 American College of Sports Medicine