Activity Recognition from Accelerometer Data

Nishkam RaviNikhil DandekarPreetham MysoreMichael L. Littman

article2005AAAI1,782 citations

Demonstrates that everyday physical activities can be accurately identified using a single pelvic-worn triaxial accelerometer combined with meta-level classification techniques like Plurality Voting.

Listen

Context and problem: Seamless ubiquitous computing and health monitoring require devices to accurately understand what a user is doing without requiring manual human input. While previous research demonstrated high recognition accuracy by placing multiple sensors across a subject's body, practical adoption requires less intrusive, low-cost configurations. Evaluating whether a single wearable sensor can reliably classify human actions under realistic conditions is therefore an essential practical challenge.

Objective: The article set out to evaluate how accurately human activities can be identified using data from a single triaxial accelerometer worn near the pelvis, while comparing the effectiveness of various base-level and combined meta-level machine learning classifiers.

Approach: The authors collected motion data from two subjects performing eight daily activities—including walking, running, stair climbing, vacuuming, and brushing teeth—over multiple days using a single pelvic-worn sensor sampling at 50 Hz. From the raw motion signals, they extracted four standard statistical features (mean, standard deviation, energy, and correlation) across overlapping time windows. The authors evaluated five individual base classifiers and several meta-classifier ensembles across four testing scenarios: randomly split mixed data from single and multiple subjects, as well as more challenging non-random setups evaluating cross-day and cross-subject generalization.

Key findings: First, meta-level classifier combinations systematically outperformed individual algorithms, with Plurality Voting achieving the most robust and consistent performance across all test environments (reaching up to 99.82% accuracy in mixed data and maintaining a leading 0.78 performance correlation between randomized and non-randomized environments). Second, model accuracy drops significantly in realistic non-random environments; while mixed data yielded accuracies between 91% and 99.8%, evaluating models across different days or different subjects lowered accuracy to roughly 46% to 90.6%. Third, feature ablation showed that standard deviation, correlation, and mean are critical for accuracy, whereas the energy feature is least significant and can be omitted without meaningful performance degradation. Finally, upper-body and fine-motor activities (such as brushing teeth) or directionally similar actions (climbing up versus down stairs) proved substantially harder to distinguish using a single pelvic sensor.

Implications and interpretation: These findings demonstrate that a single pelvic accelerometer combined with ensemble voting offers a viable, computationally efficient approach for general activity tracking. However, stakeholders should recognize the performance gap when deploying models on new users or different days without individualized calibration. Furthermore, because energy features contribute minimally, developers can reduce processing overhead and save battery life on mobile devices by removing complex frequency-domain calculations without hurting classification quality.

Recommendations and next steps: Product teams and researchers should adopt Plurality Voting when implementing wearable activity models to ensure balanced performance across diverse users. To reduce power consumption on battery-constrained devices, developers should prioritize standard deviation and axis correlation while dropping FFT-based energy calculations. Future work should explore activity ontologies to better distinguish ambiguous movements, and evaluate whether brief actions (such as swiping an access card) can be effectively recognized.

Limitations and confidence: The study provides high confidence regarding the relative advantages of Plurality Voting over other classifiers on pelvic sensor data. However, overall confidence in universal real-world performance is bounded by a very small sample size of two subjects and the recognized limitation that a pelvic-mounted sensor cannot reliably capture localized upper-body movements.

Rutgers University accelerometer data triaxial accelerometer gyroscopes Classi Ers Data Activity

Ravi et al (2005).pdf

Paper: Popular Ensemble Methods: An Empirical Study, David Opitz et al. (1999). Provides the foundational empirical framework for evaluating ensemble and meta-level classification techniques that the source applies to accelerometer data.
Paper: An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants, E. Bauer et al. (1999). Establishes how voting, bagging, and boosting mechanisms operate across base classifiers, directly underpinning the plurality voting and meta-classification explored in the source.
Paper: Toward Machine Emotional Intelligence: Analysis of Affective Physiological State, Rosalind W. Picard et al. (2001). Demonstrates early methods for feature extraction and pattern recognition from wearable sensor data to infer user state within context awareness.
Paper: Solving Multiclass Learning Problems via Error-Correcting Output Codes, Thomas G. Dietterich et al. (1994). Introduces foundational multiclass decomposition strategies that inform the multi-activity classification formulations evaluated in the source.

Paper: Deep Learning for Sensor-based Activity Recognition: A Survey, Jindong Wang et al. (2017). Surveys the evolution from classical shallow classifiers on sensor data, as studied in the source, to end-to-end deep learning architectures for sensor-based activity recognition.
Paper: Deep learning for time series classification: a review, Hassan Ismail Fawaz et al. (2018). Extends the study of time-series sensor data classification by evaluating modern deep neural networks against traditional classifier baselines.
Paper: Learning from Time-Changing Data with Adaptive Windowing, Albert Bifet et al. (2007). Builds upon static-window sensor stream classification by introducing adaptive windowing to handle real-time temporal concept drift.
Paper: An empirical comparison of supervised learning algorithms, R. Caruana et al. (2006). Expands on the comparison of base-level and ensemble classifiers by benchmarking a wider array of supervised learning algorithms across multiple evaluation metrics.
Paper: Context-aware recommender systems, Gediminas Adomavicius et al. (2008). Applies user activity and sensor-derived context awareness toward downstream ubiquitous computing systems and personalized recommendation engines.

Cover for Activity Recognition from Accelerometer Data

Abstract

Activity recognition fits within the bigger framework of context awareness. In this paper, we report on our efforts to recognize user activity from accelerometer data. Activity recognition is formulated as a classification problem. Performance of base-level classifiers and meta-level classifiers is compared. Plurality Voting is found to perform consistently well across different settings.

Topics

accelerometer data triaxial accelerometer gyroscopes

Keywords

Classi Ers Data Activity Accelerometer Activity Recognition Accelerometer Data Base-level Classi Ers Plurality Voting Activities Meta-level Classi Ers

Introduction
Data Collection
Feature extraction
Data Interpretation
Results
Conclusions and Future work
Acknowledgments
References

Knowls

Knowl 1 — Classifier Performance Across IID and Non-IID Accelerometer Evaluation Settings

data/table

Activity recognition accuracy was evaluated across five base classifiers (Naive Bayes, Support Vector Machines, $k$ -Nearest Neighbors, Decision Tables, Decision Trees / C4.5) and five meta-classification techniques (Boosting, Bagging, Plurality Voting, Stacking with Ordinary Decision Trees [ODTs], Stacking with Meta Decision Trees [MDTs]) under four experimental data distribution regimes:

Setting 1: Single-subject data collected over multiple days, mixed and evaluated via 10-fold cross-validation (IID).
Setting 2: Multi-subject data collected over multiple days, mixed and evaluated via 10-fold cross-validation (IID).
Setting 3: Single-subject data trained on day 1 and tested on day 2 (non-IID across days).
Setting 4: Data trained on subject 1 (day 1) and tested on subject 2 (day 2) (non-IID across subjects and days).

Classifier	Setting 1 (%)	Setting 2 (%)	Setting 3 (%)	Setting 4 (%)
Naive Bayes (NB)	98.86	96.69	89.96	64.00
Boosted NB	98.86	98.71	89.96	64.00
Bagged NB	98.58	96.88	90.39	59.33
SVM	98.15	98.16	68.78	63.00
Boosted SVM	99.43	98.16	67.90	73.33
Bagged SVM	98.15	98.53	68.78	60.00
kNN	98.15	99.26	72.93	49.67
Boosted kNN	99.15	99.26	72.93	49.67
Bagged kNN	99.15	99.26	70.52	46.67
Decision Table (DT)	92.45	91.91	55.68	46.33
Boosted DT	97.86	98.53	55.68	46.33
Bagged DT	93.30	94.85	55.90	46.67
Decision Tree (DTr)	97.29	98.53	77.95	57.00
Boosted DTr	98.15	98.35	77.95	57.00
Bagged DTr	97.29	95.22	78.82	63.33
Plurality Voting	99.57	99.82	90.61	65.33
Stacking (MDTs)	99.00	99.26	89.96	64.00
Stacking (ODTs)	98.86	98.35	84.50	64.00

Meta-level classifiers generally outperform base-level classifiers. Plurality Voting achieves the highest accuracy in Settings 1, 2, and 3, and achieves the highest performance correlation (0.78) between IID (Settings 1 and 2) and non-IID (Settings 3 and 4) evaluations. Boosted SVM achieves the highest accuracy in the cross-subject non-IID Setting 4 (73.33%).

Knowl 2 — Error Correlation Criterion Explaining Plurality Voting Superiority over Stacking
theoretical result
The relative advantage of Plurality Voting over Meta Decision Tree (MDT) and Ordinary Decision Tree (ODT) stacking on accelerometer data is governed by the base classifier error correlation. The error correlation $\phi(C_i, C_j)$ between two classifiers $C_i$ and $C_j$ is defined as the conditional probability that both classifiers make the identical error given that at least one of them makes an error:

$\phi(C_i, C_j) = p(C_i(x) = C_j(x) \mid C_i(x) \neq c(x) \lor C_j(x) \neq c(x))$

where $C_i(x)$ and $C_j(x)$ are the predicted classes for instance $x$ , and $c(x)$ is the true class.

Stacking methods (ODTs and MDTs) typically outperform Plurality Voting when base classifier error diversity is high (error correlation $< 50\%$ ). When error correlation exceeds approximately $50\%$ , Plurality Voting outperforms stacking. Across all four accelerometer evaluation settings, the empirical average pairwise error correlation among the base-level classifiers is approximately $52\%$ , explaining why Plurality Voting consistently outperforms MDT and ODT stacking.
Knowl 3 — Triaxial Accelerometer Signal Windowing and Feature Extraction
model/method
Raw triaxial accelerometer signals ( $x, y, z$ axes sampled at 50 Hz) are segmented into sliding windows of $N = 256$ samples (representing $5.12\text{ s}$ of data) with a $50\%$ overlap (128 samples between successive windows). For each window, four statistical and spectral features are computed across all three axes, generating a 12-dimensional feature vector:
1. Mean: The direct-current (DC) component representing average acceleration along the axis over the window.
2. Standard Deviation: Captures the spread and amplitude range of movement dynamics along the axis.
3. Energy: Captures signal periodicity in the frequency domain, defined as the normalized sum of squared discrete Fast Fourier Transform (FFT) component magnitudes:
$\text{Energy} = \frac{\sum_{i=1}^{|w|} |x_i|^2}{|w|}$

where $x_1, x_2, \dots, x_{|w|}$ are the discrete FFT component magnitudes of the window $w$ , and $|w| = 256$ . 4. Correlation: Computed between each pair of axes $(x,y)$ , $(y,z)$ , and $(x,z)$ to capture multi-dimensional translation dynamics:

$\operatorname{corr}(x, y) = \frac{\operatorname{cov}(x, y)}{\sigma_x \sigma_y}$

where $\operatorname{cov}(x, y)$ is the covariance between axes $x$ and $y$ , and $\sigma_x, \sigma_y$ are their respective standard deviations.
Knowl 4 — Four-Setting Experimental Protocol for Evaluating Sensor Generalization
experimental setup
The experimental setup captures user activities using a single Crossbow CDXL04M3 triaxial accelerometer (sensing range up to $\pm 4\text{G}$ , tolerance within $2\%$ ) mounted on a hoarder board sampling at 50 Hz, worn near the pelvic region. Sensor readings are transmitted wirelessly via Bluetooth to an HP iPAQ.

Data was gathered for eight physical activities:
- Standing
- Walking
- Running
- Climbing up stairs
- Climbing down stairs
- Sit-ups
- Vacuuming
- Brushing teeth
Activities were performed by two subjects across multiple days. Labels were generated semi-automatically via stop-watch logs, and data within 10 seconds of recorded start and stop times were removed to avoid transition artifacts. Classifiers are tested across four settings to distinguish IID behavior from temporal and cross-subject non-IID shifts:
- Setting 1 (IID): Single subject, mixed across days (10-fold CV).
- Setting 2 (IID): Two subjects, mixed across days (10-fold CV).
- Setting 3 (non-IID): Single subject, trained on day 1 data, tested on day 2 data.
- Setting 4 (non-IID): Trained on subject 1 (day 1), tested on subject 2 (day 2).

Knowl 5 — Feature Ablation on Accelerometer Activity Recognition

data/table

To assess attribute relevance, classifiers were evaluated on Setting 2 data (multi-subject, mixed-day cross-validation) by dropping one feature type across all three axes at a time.

Standard deviation is the most critical feature (its removal increases average misclassifications from 14.05 to 32.44), followed by inter-axis correlation (28.38) and mean acceleration (21.83). Energy is the least significant attribute, as its omission causes negligible degradation in classification accuracy (14.72 vs 14.05 misclassifications).

Knowl 6 — Activity Confusion Patterns in Pelvic-Worn Accelerometer Sensing

data/table

Analysis of the representative confusion matrix for Setting 3 (single-subject, trained on day 1, tested on day 2) highlights specific failure modes of a pelvic-worn triaxial accelerometer.

Activity	Standing	Walking	Running	Stairs Up	Stairs Down	Vacuuming	Brushing	Situps
Standing	63	0	0	0	0	0	0	0
Walking	0	44	0	1	0	0	0	0
Running	0	0	17	16	20	0	0	0
Stairs Up	0	0	0	9	12	0	0	0
Stairs Down	0	0	0	19	0	0	0	0
Vacuuming	0	0	0	0	0	45	0	0
Brushing	18	0	0	0	0	15	0	0
Situps	0	0	0	0	0	7	0	24

Two main confusion patterns occur:

Ascending vs. Descending Stairs: Climbing stairs up and climbing stairs down are frequently confused with each other (e.g., 12 stairs-up instances classified as stairs-down, and 19 stairs-down instances classified as stairs-up) and occasionally with running.
Localized Upper-Body Activities: Activities confined largely to upper-body or arm/mouth movements (such as brushing teeth) generate minimal distinctive pelvic acceleration, causing brushing to be frequently misclassified as standing (18 instances) or vacuuming (15 instances).

Knowl 7 — Probability Distribution Summation for Plurality Voting Meta-Classification
model/method
When base-level classifiers output class probability distributions (CPDs) rather than hard class labels, Plurality Voting aggregates predictions by averaging the predicted probability distribution vectors over the base ensemble:

$P_{ML}(x) = \frac{1}{|C|} \sum_{c \in C} P_c(x)$

where $C$ is the set of base-level classifiers (such as Naive Bayes, Support Vector Machines, $k$ -Nearest Neighbors, Decision Tables, and C4.5 Decision Trees), $P_c(x)$ is the class probability distribution vector produced for input instance $x$ by classifier $c$ , and $P_{ML}(x)$ is the resulting meta-level class probability distribution. The final predicted class is the class maximizing $P_{ML}(x)$ .

Coverage note — None was omitted; all key experimental protocols, data tables, feature extraction formulas, error-correlation analysis, and empirical conclusions are represented.

References

1.Bao, L., and Intille, S. S. 2004. Activity recognition from user-annotated acceleration data. In Proceceedings of the 2nd International Conference on Pervasive Computing, 1–17.
2.Breiman, L. 1996. Bagging predictors. Machine Learning 123–140.
3.Bussmann, J.; Martens, W.; Tulen, J.; Schasfoort, F.; van den Bergemons, H.; and H.J.Stam. 2001. Measuring daily behavior using ambulatory accelerometry: the activity monitor. Behavior Research Methods, Instruments, and Computers 349–356.
4.DeVaul, R., and Dunn, S. 2001. Real-Time Motion Classification for Wearable Computing Applications. Technical report, MIT Media Laboratory.
5.Dzeroski, S., and Zenko, B. 2004. Is combining classifiers with stacking better than selecting the best one? Machine Learning 255–273.
6.Foerster, F.; Smeja, M.; and Fahrenberg, J. 1999. Detection of posture and motion by accelerometry: a validation in ambulatory monitoring. Computers in Human Behavior 571–583.
7.Freund, Y., and Schapire, R. E. 1996. Experiments with a new boosting algorithm. In International Conference on Machine Learning, 148–156.
8.Gama, J., and Brazdil, P. 2000. Cascade generalization. Machine Learning 315–343.
9.Harter, A., and Hopper, A. 1994. A distributed location system for the active office. IEEE Network 8(1).
10.Lee, S., and K.Mase. 2002. Activity and location recognition using wearable sensors. IEEE Pervasive Computing 24–32.
11.Makikawa, M.; Kurata, S.; Higa, Y.; Araki, Y.; and Tokue, R. 2001. Ambulatory monitoring of behavior in daily life by accelerometers set at both-near-sides of the joint. In Proceedings of MedInfo, 840–843.
12.Meir, R., and Ratsch, G. 2003. An introduction to boosting and leveraging. 118–183.
13.Priyantha, N. B.; Chakraborty, A.; and Balakrishnan, H. 2000. The cricket location-support system. In Mobile Computing and Networking, 32–43.
14.Randell, C., and Muller, H. 2000. Context awareness by analysing accelerometer data. In MacIntyre, B., and Iannucci, B., eds., The Fourth International Symposium on Wearable Computers, 175–176. IEEE Computer Society.
15.Seewald, A. K. 2002. How to make stacking better and faster while also taking care of an unknown weakness. In Proceedings of the Nineteenth International Conference on Machine Learning, 554–561. Morgan Kaufmann Publishers Inc.
16.Todorovski, L., and Dzeroski, S. 2003. Combining classifiers with meta decision trees. Machine Learning 223–249.
17.Want, R.; Hopper, A.; Falcao, V.; and Gibbons, J. 1992. The active badge location system. Technical Report 92.1, ORL, 24a Trumpington Street, Cambridge CB2 1QA.
18.Witten, I., and Frank, E. 1999. Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. Morgan Kauffman.
19.Wolpert, D. H. 1992. Stacked generalization. Neural Networks 241–259.

Citation

MLA

Ravi, N., et al. “Activity Recognition from Accelerometer Data”. Innovative Applications of Artificial Intelligence, 2005, pp. 1541–46, http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.557.6791.

APA

Ravi, N., Dandekar, N., Mysore, P., & Littman, M. L. (2005). Activity recognition from accelerometer data. Innovative Applications of Artificial Intelligence, 1541–1546. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.557.6791

Chicago

Ravi, N., N. Dandekar, P. Mysore, and M. L. Littman. 2005. “Activity Recognition from Accelerometer Data”. Innovative Applications of Artificial Intelligence, 1541–46. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.557.6791.

Harvard

Ravi, N. et al. (2005) “Activity recognition from accelerometer data”, Innovative Applications of Artificial Intelligence, pp. 1541–1546. Available at: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.557.6791.

Vancouver

1. Ravi N, Dandekar N, Mysore P, Littman ML (2005) Activity recognition from accelerometer data. Innovative Applications of Artificial Intelligence 1541–1546

BibTeX

@article{ravi2005activity,
  title = {Activity recognition from accelerometer data},
  author = {Ravi, Nishkam and Dandekar, Nikhil and Mysore, Preetham and Littman, Michael L.},
  year = {2005},
  journal = {Innovative Applications of Artificial Intelligence},
  pages = {1541-1546},
  url = {http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.557.6791}
}

Metadata:DOI registry

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Attribute Dropped	Average Number of Misclassifications
Drop None	14.05
Drop Mean	21.83
Drop Standard Deviation	32.44
Drop Energy	14.72
Drop Correlation	28.38

Activity Recognition from Accelerometer Data

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Classifier Performance Across IID and Non-IID Accelerometer Evaluation Settings

Knowl 2 — Error Correlation Criterion Explaining Plurality Voting Superiority over Stacking

Knowl 3 — Triaxial Accelerometer Signal Windowing and Feature Extraction

Knowl 4 — Four-Setting Experimental Protocol for Evaluating Sensor Generalization

Knowl 5 — Feature Ablation on Accelerometer Activity Recognition

Knowl 6 — Activity Confusion Patterns in Pelvic-Worn Accelerometer Sensing

Knowl 7 — Probability Distribution Summation for Plurality Voting Meta-Classification

References

Citation

BibTeX

Access the Paper