Feature extraction for robust physical activity recognition
Open accessAbstract
This paper presents the development of a Human Activity Recognition (HAR) system that uses a network of nine inertial measurement units situated in different body parts. Every unit provides 3D (3-dimension) acceleration, 3D angular velocity, 3D magnetic field orientation, and 4D quaternions. This system identifies 33 different physical activities (walking, running, cycling, lateral elevation of arms, etc.). The system is composed of two main modules: a feature extractor for obtaining the most relevant characteristics from the inertial signals every second, and a machine learning algorithm for classifying between the different activities. This paper focuses on the feature extractor module, evaluating several types of features and proposing different normalization approaches. This paper also analyses the performance of every sensor included in the inertial measurement units. The main experiments have been done using a public available dataset named REALDISP Activity Recognition dataset. This dataset includes recordings from 17 subjects performing 33 different activities in three different scenarios. Final results demonstrate that the proposed HAR system significantly improves the classification accuracy compared to previous works on this dataset. For the best configuration, the system accuracy is 99.1%. This system has been also evaluated with the OPPORTUNITY dataset obtaining competitive results.
AI snapshot
Written from the paper’s full text.
TL;DR
This paper develops a wearable-sensor human activity recognition system using nine inertial measurement units and a feature-extraction plus Random Forest pipeline to classify 33 physical activities. On the REALDISP dataset, the best configuration reaches 99.1% accuracy and 0.991 F-measure, improving on prior work and showing robustness to sensor placement and inter-user variability.
Key findings
- The best REALDISP configuration achieves 99.1% accuracy and 0.991 F-measure with subject-wise evaluation and Null-activity included; with Null-activity truncated under ideal placement, accuracy is 99.4% versus the baseline 97.0%.
- Subject-wise evaluation is substantially harder than random-partitioning: for accelerometer/time-based features, accuracy drops from 99.1% to 95.5%, a 3.4% difference.
- Magnetometer signals perform best among single sensor types (96.3% accuracy, 0.952 F-measure), while quaternion performs worst (93.0%, 0.924); combining all sensors improves accuracy to 97.0% before normalization.
- Time-domain features outperform frequency-domain features and their combination; vector normalization on signal data (SIGN 5) is best, e.g. magnetometer reaches 97.9% accuracy and 0.976 F-measure.
- The system is more robust across placement scenarios: self-placement reaches 98.9% versus baseline 88.4%, and mutual-placement results are 87.9%, 93.5%, 96.4%, and 83.2% for 4–7 misplaced sensors versus baselines of 71.2%, 71.6%, 77.2%, and 68.0%.
Methods
- Design: a two-module HAR system with feature extraction from 3 s sliding windows (66% overlap, 150 samples per frame) and a Random Forest classifier; 9 inertial measurement units provide 117 inertial signals, and feature vectors contain 4086 features.
- Data: main experiments use the REALDISP Activity Recognition dataset with 17 subjects (7 female, 10 male, ages 22–37), 33 activities, 9 on-body units, 13 inertial signals per unit, sampled at 50 Hz; additional validation uses the OPPORTUNITY dataset.
- Analysis: compares tenfold random-partitioning and subject-wise cross-validation, evaluates sensor types, time- versus frequency-domain features, six normalization methods, placement scenarios, and OPPORTUNITY high-level locomotion and mid-level gesture tasks; significance is assessed with a confidence interval lower than 0.5%.
Limitations
- The paper states that in the mutual-placement scenario only 3 of 17 volunteers were recorded, so training data are very small (2 of 3 subjects for training), and mutual-placement results are considerably lower than ideal-placement results.
- The paper states that the amount of training data is an important factor, as training with ideal-placement data improves mutual-placement accuracy, e.g. mutual4 from 87.9% to 99.0%.
- Evident from the text: the OPPORTUNITY evaluation truncates object and ambient sensors and uses only body-worn sensors; feature extraction is computationally heavy, needing 12 min for a whole session of about 18 min of exercise, and training needs less than 10 min.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 16 cited sources checked · 0 supported · 0 flagged
- Findingnone given
The proposed HAR system identifies 33 different physical activities.
“This system identifies 33 different physical activities (walking, running, cycling, lateral elevation of arms, etc.).”
Quote in the paperNot a substantial claim - Findingown data
The proposed HAR system significantly improves classification accuracy compared to previous works on the REALDISP dataset.
“Final results demonstrate that the proposed HAR system significantly improves the classification accuracy compared to previous works on this dataset.”
- Also stated in: § Introduction · boosted
“The results obtained in this paper significantly improve the accuracy obtained in previous works on the same dataset.”
Quote in the paperSubstantial claim - Also stated in: § Introduction · boosted
- Findingown data
For the best configuration, the system accuracy is 99.1%.
“For the best configuration, the system accuracy is 99.1%.”
Quote in the paperSubstantial claim - Findingown data
Temporal features always beat frequency features and their combination for all three sensor types.
“From the results shown in Fig. 2, it is obvious that the temporal features always beat the frequency features and their combination in the cases of all three sensor types.”
Quote in the paperSubstantial claim - Findingown data
In the ideal placement scenario, after introducing the signal normalization method and considering all sensors, the proposed system's accuracy is 99.4%.
“After introducing the signal normalization method and considering all sensors, the system accuracy goes to 99.4%, a 2.4% improvement compared to the original paper (97%).”
Quote in the paperSubstantial claim
- Findingown data
Including all sensors yields even higher system accuracy: 97.0%.
“By including all sensors, we obtain even higher system accuracy: 97.0%.”
Quote in the paperSubstantial claim - Findingown data
Vector normalization applied directly on signal data outperforms all other normalization methods.
“The results show that the vector normalization method (number 5) applied directly on the signal data outperforms all other methods.”
Quote in the paperSubstantial claim - Findingown data
When the Null-activity is included, the proposed system's accuracy is 99.1%.
“our system still maintains a high performance when the Null-activity is included: the system only loses 0.3% accuracy (from 99.4 to 99.1%) showing a significant improvement (2.1%) respect to the baseline system (97%).”
Quote in the paperSubstantial claim - Findingown data
In the self-placement scenario, the original paper's accuracy drops 8.6% compared to ideal placement (from 97.0% to 88.4%).
“Regarding the self-placement scenario, this table shows 8.6% accuracy drop compared to the ideal-placement in the original paper (from 97.0 to 88.4%), but with our system, this reduction is lower than 0.5% (from 99.1 to 98.9%).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
- Findingown data
Including the Null-activity reduces the proposed system's accuracy by 0.3% (from 99.4% to 99.1%).
“our system still maintains a high performance when the Null-activity is included: the system only loses 0.3% accuracy (from 99.4 to 99.1%) showing a significant improvement (2.1%) respect to the baseline system (97%).”
Quote in the paperSubstantial claim - Findingown data
In the self-placement scenario, the proposed system's accuracy reduction compared to ideal placement is lower than 0.5% (from 99.1% to 98.9%).
“Regarding the self-placement scenario, this table shows 8.6% accuracy drop compared to the ideal-placement in the original paper (from 97.0 to 88.4%), but with our system, this reduction is lower than 0.5% (from 99.1 to 98.9%).”
Quote in the paperSubstantial claim - Findingown data
Including the Null-activity, the proposed system shows a 2.1% improvement over the baseline system's 97% accuracy.
“our system still maintains a high performance when the Null-activity is included: the system only loses 0.3% accuracy (from 99.4 to 99.1%) showing a significant improvement (2.1%) respect to the baseline system (97%).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
- Findingown data
The proposed system's accuracy in ideal placement is a 2.4% improvement over the original paper's 97% baseline.
“After introducing the signal normalization method and considering all sensors, the system accuracy goes to 99.4%, a 2.4% improvement compared to the original paper (97%).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
- Findingown data
In the mutual-placement scenario, the results are considerably low in both the baseline and the proposed system.
“For the mutual-placement scenario, the results are considerably low in both works (baseline and this paper) but the degradation obtained with the system proposed in this paper is considerably smaller compared to the baseline system: our system shows a better robustness.”
Quote in the paperNot a substantial claim - Findingreasoning
In the self-placement scenario, the proposed system improves accuracy by more than 10% over the original paper (from 88.4% to 98.9%).
“When comparing these results to the original paper, there is a big improvement (more than 10%, from 88.4 to 98.9%) in the self-placement scenario.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
- Findingown data
Final results show that the proposed system largely improves the performance compared to previous works on the same dataset [24].
“Final results have exhibited that the proposed system largely improves the performance compared to previous works on the same dataset [24].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
- Findingown data
For the high-level locomotion task on the OPPORTUNITY dataset, the proposed system obtains the best results for all subjects when the Null class is not considered.
“For the high-level locomotion task (Table 7), the system proposed in this paper obtains the best results for all subjects when the Null class is not considered (the 4 last columns).”
Quote in the paperNot a substantial claim - Findingnone given
The time-based features have provided better results compared to the frequency-based ones.
“The time-based features have provided better results compared to the frequency-based ones.”
Quote in the paperSubstantial claim - Findingnone given
The degradation in mutual-placement is different depending on the number of mis-displacements (4, 5, 6, and 7).
“This degradation is different depending on the number of mis-displacements (4, 5, 6, and 7).”
Quote in the paperNot a substantial claim - Findingnone given
The best results have been obtained when combining the information from all the sensors.
“The best results have been obtained when combining the information from all the sensors.”
Quote in the paperSubstantial claim - Findingreasoning
The improvement from combining all sensors is significant.
“In this case, the improvement is significant.”
Quote in the paperSubstantial claim - Findingown data
For the mid-level gesture task on the OPPORTUNITY dataset, the proposed system obtains the best results for all subjects except S4 when the Null class is included.
“For the mid-level gesture task (Table 8), the system proposed in this paper obtains the best results for all subjects except S4 when the Null class is included (the 4 first”
Quote in the paperNot a substantial claim - Findingown data
When trained with ideal datasets and tested with mutual datasets, the system reaches very good accuracy even though training and testing sets come from different placement scenarios.
“The results show that when being trained with ideal datasets and tested with mutual datasets, the system reaches a very good accuracy though the training and testing sets come from different placement scenarios.”
Quote in the paperSubstantial claim - Findingown data
For mutual4, the accuracy goes from 87.9% to 99.0% when training with ideal instead of mutual data.
“For example, for mutual4, the accuracy goes from 87.9 to 99.0% (the first row).”
Quote in the paperSubstantial claim - Findingown data
Under the best configuration, the proposed system's F-measure is 0.991.
“Under the best configuration, the accuracy reaches 99.1% and F-measure 0.991.”
Quote in the paperSubstantial claim - Findingown data
In the cross-dataset experiment, the amount of data available in ideal-placement and self-placement scenarios is the same.
“In this case, the amount of data available in ideal-placement and self-placement scenarios is the same.”
Quote in the paperSubstantial claim - Findingown data
The proposed system has been evaluated with the OPPORTUNITY dataset, demonstrating competitive results compared to previous work [31] in high-level locomotion and mid-level gesture classification for home care monitoring.
“The proposed system has been also evaluated with another public dataset (OPPOR-TUNITY dataset) demonstrating competitive results (compared to previous work [31]) in two main tasks for home care monitoring: high-level locomotion and mid-level gesture classification.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [31] Chavarriaga R, Sagha H, Calatroni A, Digumarti ST, Tröster G, Millán JDR, Roggen D (2013) The OPPORTUNITY challenge: a benchmark database for on-body sensor-based activity recognition. Pattern Recogn Lett 34:2033–2042 source
Not checked in the free run.
- Findingnone given
The magnetometer signals have provided better discrimination capability than other sensor types.
“Regarding the type of sensor, the magnetometer signals have provided better discrimination capability.”
Quote in the paperSubstantial claim - Findingown data
There is no significant difference in accuracy when testing with the self-placement dataset and training with ideal or self placement (99.1% vs. 98.9%, difference lower than the 0.5% confidence interval).
“As Table 6 shows, there is not a significant difference on the accuracy when testing with self-placement dataset and training with ideal or self placement (99.1 vs. 98.9%, difference lower than the confidence interval 0.5%).”
Quote in the paperSubstantial claim - Findingown data
For the high-level locomotion task on the OPPORTUNITY dataset, when including the Null class, the proposed system obtains the best results for all subjects except S3.
“When including the Null class (the 4 first columns), we obtain the best results for all subjects except S3.”
Quote in the paperNot a substantial claim - Interpretationown data
The better performance of random-partitioning evaluation is because training and testing subsets could contain information from the same subject, allowing the algorithm to learn subject-dependent characteristics.
“This result supports the hypothesis stated in the previous section: in the random-part evaluation, training and testing subsets could contain information from a same subject and this characteristic produces better classification results.”
Quote in the paperSubstantial claim - Interpretationnone given
The proposed HAR system reports the best HAR results using the REALDISP dataset as far as the authors know.
“As far as the authors know, this work reports the best HAR results using this dataset.”
Quote in the paperNot a substantial claim - Premisecitation
The REALDISP dataset includes 13 inertial signals obtained from 9 on-body inertial measurement units.
“These recording include 13 inertial signals obtained from 9 on-body inertial measurement units located on different body parts.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
- Premisecitation
The REALDISP dataset includes sensor displacement such as rotation and mis-positioning from the corresponding body part.
“The most significant characteristic of this dataset is the introduction of sensor displacement, such as rotation and mis-positioning from the corresponding body part.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
- Interpretationown data
Using signals from magnetometer and time-based features is the best system configuration.
“As a conclusion, it is clear that using the signals from magnetometer and the timebased features is currently the best system configuration.”
Quote in the paperSubstantial claim - Interpretationown data
The new feature extraction module shows very good robustness against different sensor placements.
“The new feature extraction module shows a very good robustness against different sensor placements.”
Quote in the paperSubstantial claim - Interpretationreasoning
The low degradation when including Null-activity is made possible by the large number of features extracted and the suitable normalization method proposed in this paper.
“Such low degradation is made possible due to the large number of features extracted and the suitable normalization method proposed in this paper.”
Quote in the paperSubstantial claim - Interpretationown data
Thanks to the new feature extraction module and the normalization strategy, the system has shown strong robustness when facing the Null-activity and different placement scenarios.
“Thanks to the new feature extraction module and the normalization strategy, the system has shown strong robustness when facing the Null-activity and different placement scenarios, two vital aspects for real applications.”
Quote in the paperSubstantial claim - Interpretationown data
The proposed system is a competitive solution for home care monitoring applications.
“In conclusion, the proposed system is also a competitive solution for home care monitoring applications.”
Quote in the paperSubstantial claim - Interpretationown data
These results support the hypothesis that the amount of data for training is an important factor in the system performance.
“These results support the hypothesis that the amount of data for training is an important factor in the system performance.”
Quote in the paperSubstantial claim
- May combine two claims. 10
- Classified as probably not a substantial claim. 6
May combine two claims.
Findingown dataThe random-partitioning evaluation method yields significantly better accuracy than subject-wise evaluation (99.1% vs 95.5%, difference 3.4% > CI 0.5%).
“From the results, we can clearly see that the result given by the random-partitioning method is significantly better than the subject-wise method: the accuracy (Acc%) difference is 3.4% (99.1-95.5%)higher than the confidence interval, 0.5%.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataThe 3D magnetometer works best among the four sensor types and the quaternion sensor performs worst.
“From the results, we can clearly see that the 3D magnetometer works best among the four types of sensor and the quaternion sensor performs the worst.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone givenThe system evaluated with the OPPORTUNITY dataset obtains competitive results.
“This system has been also evaluated with the OPPORTUNITY dataset obtaining competitive results.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataRandom Forest works better than J48 decision tree in this circumstance, defeating it by nearly 10% in accuracy in preliminary experiments.
“In our preliminary experiments, it defeats the J48 decision tree algorithm by nearly 10% in accuracy.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningThe 2.4% improvement over the original paper is statistically significant with 95% probability because it is higher than the 0.5% confidence interval.
“This improvement is higher than the confidence interval (0.5%) so the difference is statistically significant with a 95% of probability.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
PremisecitationThe REALDISP dataset includes recordings from 17 subjects, seven females and ten males, ages 22 to 37.
“This dataset includes recordings from 17 subjects, seven females and ten males, with ages ranging from 22 to 37 years old.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
May combine two claims.
Premiseown dataThe REALDISP dataset includes recordings from 17 subjects performing 33 different activities in three placement scenarios.
“This dataset includes recordings from 17 subjects performing 33 different activities in three different scenarios.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationThe REALDISP dataset includes three placement scenarios: Ideal-placement, Self-placement, and Mutual-placement.
“The dataset includes recordings in three placement scenarios:”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationThe REALDISP dataset samples 3D linear acceleration, 3D angular velocity, 3D magnetic field orientation and 4D quaternions every 20 ms (50 Hz).
“Using these sensors, a 3D (3-dimension) linear acceleration, a 3D angular velocity, a 3D magnetic field orientation and a 4D quaternions are sampled every 20 ms (50 Hz sample-rate).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
PremisecitationThe REALDISP dataset includes a Null-activity label assigned to other activities and transitions between activities.
“This dataset also includes a Null-activity.This label has been assigned to other activities (different from the 33 activities considered in this study), and also, the transitions between activities.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
May combine two claims.
Study design and conductreasoningIn this paper, for all experiments, the confidence interval is lower than 0.5%, so any difference higher than 0.5% is significant with 95% probability.
“In this paper, for all the experiments, the confidence interval is lower than 0.5%, so any difference higher than this value, this difference can be considered as significant with a 95% a probability.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [29] Cox DR, Hinkley DV (1974) Theoretical statistics. Chapman & Hall, London source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationThe REALDISP dataset includes 33 physical activities including warm up, fitness and cool down activities.
“The experiment consisted in performing a complete set of exercises: 33 physical activities, including warm up, fitness and cool down activities (walking, jogging, cycling, jumping, etc.).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
May combine two claims.
InterpretationcitationIn the original paper, the evaluation method was random-partitioning, and the baseline accuracy would be even lower when using subject-wise cross-validation.
“It is important to notice that, in the original paper, the evaluation method was random-partitioning and, based on the results presented in”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] Banos O, Toth MA, Damas M, Pomares H, Rojas I (2014) Dealing with the effects of sensor displacement in wearable activity recognition. Sensors 14:9995–10023 source
Not checked in the free run.
May combine two claims.
Interpretationown dataIn the mutual-placement scenario, the degradation obtained with the proposed system is considerably smaller compared to the baseline system, showing better robustness.
“For the mutual-placement scenario, the results are considerably low in both works (baseline and this paper) but the degradation obtained with the system proposed in this paper is considerably smaller compared to the baseline system: our system shows a better robustness.”
Quote in the paperSubstantial claim
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
A Random Forest algorithm was selected as classifier for the ARC because Zhu et al. (2017) reported that the random forest classifier yielded the best results in a HAR problem with a high dimensional feature space.
“We have selected a Random Forest (RF) algorithm as classifier for our ARC since Zhu et al. (2017) used several classifiers in a HAR problem with a high dimensional feature space and reported that the random forest classifier yielded the best results.”
From Advanced Sensing and Human Activity Recognition in Early Intervention and Rehabilitation of Elderly People (2020) · cites this paper as Zhu et al. (2017)
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Accuracy difference is 3.6, not 3.4 percentage points
MinorSubtracting the subject-wise accuracy of 95.5% from the random-partitioning accuracy of 99.1% gives 3.6 percentage points, not 3.4.
Reported 3.4 · recomputed 3.6
“the accuracy (Acc%) difference is 3.4% (99.1-95.5%)”
Self-placement comparison does not match Table 6
MinorThe text compares accuracies of 99.1% and 98.9% for the two training sets, but Table 6 reports 98.9% for both.
“there is not a significant difference on the accuracy when testing with self-placement dataset and training with ideal or self placement (99.1 vs. 98.9%, difference lower than the confidence interval 0.5%).”
“Self (16 subjects) Self (1 subject) 98.9 0.988 Ideal (16 subjects) Self (1 subject) 98.9 0.989”
Cites
Breiman, L. (2001) · Machine Learning · doi:10.1023/a:1010933404324 · Cited by 133,538 · see citing works
Lee, I.-M., Shiroma, E. J., Lobelo, F., Puska, P., Blair, S. N., & Katzmarzyk, P. T. (2012) · The Lancet · doi:10.1016/s0140-6736(12)61031-9 · Cited by 8,215 · see citing works
Activity Recognition from User-Annotated Acceleration Data
Bao, L., & Intille, S. S. (2004) · Lecture Notes in Computer Science · doi:10.1007/978-3-540-24646-6_1 · Cited by 3,147 · see citing works
Cox, D. R., & Hinkley, D. V. (1974) · doi:10.1007/978-1-4899-2887-0 · Cited by 3,089 · see citing works
Activity recognition using cell phone accelerometers
Kwapisz, J. R., Weiss, G. M., & Moore, S. A. (2011) · ACM SIGKDD Explorations Newsletter · doi:10.1145/1964897.1964918 · Cited by 2,912 · cited 1× · see citing works
“Kwapisz [10] compares three learning algorithms: logistic regression, J48, and multilayer perceptron.”
A survey on vision-based human action recognition
Poppe, R. (2010) · Image and Vision Computing · doi:10.1016/j.imavis.2009.11.014 · Cited by 2,242 · see citing works
Activity recognition from accelerometer data
Ravi, N., Dandekar, N., Mysore, P., & Littman, M. L. (2005) · Cited by 1,434 · see citing works
Karantonis, D. M., Narayanan, M. R., Mathie, M. J., Lovell, N. H., & Celler, B. G. (2006) · IEEE Transactions on Information Technology in Biomedicine · doi:10.1109/titb.2005.856864 · Cited by 1,239 · see citing works
Human activity recognition with smartphone sensors using deep learning neural networks
Ronao, C. A., & Cho, S.-B. (2016) · Expert Systems with Applications · doi:10.1016/j.eswa.2016.04.032 · Cited by 1,215 · see citing works
WEKA: a machine learning workbench
Holmes, G., Donkin, A., & Witten, I. H. (2002) · Australian and New Zealand Intelligent Information Systems Conference (ANZIIS) · doi:10.1109/anziis.1994.396988 · Cited by 939 · see citing works
Collecting complex activity datasets in highly rich networked sensor environments
Roggen, D., Calatroni, A., Rossi, M., Holleczek, T., Forster, K., Troster, G., Lukowicz, P., Bannach, D., Pirkl, G., Ferscha, A., et al. (2010) · 2010 Seventh International Conference on Networked Sensing Systems (INSS) · doi:10.1109/inss.2010.5573462 · Cited by 788 · see citing works
The Opportunity challenge: A benchmark database for on-body sensor-based activity recognition
Chavarriaga, R., Sagha, H., Calatroni, A., Digumarti, S. T., Tröster, G., Millán, J. D. R., & Roggen, D. (2013) · Pattern Recognition Letters · doi:10.1016/j.patrec.2012.12.014 · Cited by 742 · see citing works
Transition-Aware Human Activity Recognition Using Smartphones
Reyes-Ortiz, J.-L., Oneto, L., Samà, A., Parra, X., & Anguita, D. (2016) · Neurocomputing · doi:10.1016/j.neucom.2015.07.085 · Cited by 740 · see citing works
Stisen, A., Blunck, H., Bhattacharya, S., Prentow, T. S., Kjærgaard, M. B., Dey, A., Sonne, T., & Jensen, M. M. (2015) · Proceedings of the 13th ACM Conference on Embedded Networked Sensor Systems · doi:10.1145/2809695.2809718 · Cited by 714 · see citing works
Window Size Impact in Human Activity Recognition
Baños, O., Gálvez, J. M., Damas, M., Pomares, H., & Rojas, I. (2014) · Sensors · doi:10.3390/s140406474 · Cited by 656 · see citing works
Activity Recognition and Monitoring Using Multiple Sensors on Different Body Positions
Maurer, U., Smailagic, A., Siewiorek, D. P., & Deisher, M. E. (2006) · Wearable and Implantable Body Sensor Networks · doi:10.1109/bsn.2006.6 · Cited by 606 · see citing works
Human Activity Recognition from Accelerometer Data Using a Wearable Device
Casale, P., Pujol, O., & Radeva, P. I. (2011) · Lecture notes in computer science · doi:10.1007/978-3-642-21257-4_36 · Cited by 341 · see citing works
Toward physical activity diary
Yang, J. (2009) · doi:10.1145/1631040.1631042 · Cited by 316 · see citing works
Recognizing Workshop Activity Using Body Worn Microphones and Accelerometers
Lukowicz, P., Ward, J. A., Junker, H., Stäger, M., Tröster, G., Atrash, A., & Starner, T. E. (2004) · Lecture notes in computer science · doi:10.1007/978-3-540-24646-6_2 · Cited by 285 · see citing works
Activity Recognition and Monitoring Using Multiple Sensors on Different Body Positions
Maurer, U., Smailagic, A., Siewiorek, D. P., & Deisher, M. (2006) · doi:10.21236/ada534437 · Cited by 266 · see citing works
Unsupervised learning for human activity recognition using smartphone sensors
Kwon, Y., Kang, K., & Bae, C. (2014) · Expert Systems with Applications · doi:10.1016/j.eswa.2014.04.037 · Cited by 207 · see citing works
JUCS - Journal of Universal Computer Science
:none (2020) · CINECA IRIS Institutial research information system (University of Pisa) · doi:10.3217/jucs-019-09 · Cited by 188 · see citing works
Dealing with the Effects of Sensor Displacement in Wearable Activity Recognition
Banos, O., Toth, M., Damas, M., Pomares, H., & Rojas, I. (2014) · Sensors · doi:10.3390/s140609995 · Cited by 159 · see citing works
Jatobá, L., Grossmann, U., Kunze, C., Ottenbacher, J., & Stork, W. (2008) · Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) · doi:10.1109/iembs.2008.4650398 · Cited by 114 · see citing works
Development of intelligent multisensor surveillance systems with agents
Pavón, J., Gómez-Sanz, J. J., Fernández‐Caballero, A., et al. (2007) · Robotics and Autonomous Systems · doi:10.1016/j.robot.2007.07.009 · Cited by 106 · see citing works
A survey on using domain and contextual knowledge for human activity recognition in video streams
Onofri, L., Soda, P., Pechenizkiy, M., & Iannello, G. (2016) · Expert Systems with Applications · doi:10.1016/j.eswa.2016.06.011 · Cited by 90 · see citing works
Acoustic Event Detection and Classification
Temko, A., Nadeu, C., Macho, D., Malkin, R. E., Zieger, C., & Omologo, M. (2009) · Springer eBooks · doi:10.1007/978-1-84882-054-8_7 · Cited by 66 · see citing works
Haar-Like Filtering for Human Activity Recognition Using 3D Accelerometer
Hanai, Y., Nishimura, J., & Kuroda, T. (2009) · doi:10.1109/dsp.2009.4786008 · Cited by 65 · see citing works
Segmenting human activities based on HMMs using smartphone inertial sensors
San-Segundo, R., Lorenzo-Trueba, J., Martínez-González, B., & Pardo, J. M. (2016) · Pervasive and Mobile Computing · doi:10.1016/j.pmcj.2016.01.004 · Cited by 47 · see citing works
Self-citationA Comparative Study on Sensor Displacement Effect on Realistic Sensor Displacement Benchmark Dataset
Punchoojit, L., & Hongwarittorrn, N. (2015) · Advances in intelligent systems and computing · doi:10.1007/978-3-319-19024-2_10 · Cited by 3 · see citing works
Handling Displacement Effects in On-Body Sensor-Based Activity Recognition
Baños, O., Damas, M., Pomares, H., & Rojas, I. (2013) · Lecture notes in computer science · doi:10.1007/978-3-319-03092-0_12 · Cited by 2 · see citing works
Flagged first, then by citations
Research works printed but not linked: 2
- Anguita D, Ghio A, Oneto L, Parra X, Reyes-Ortiz JL (2013) Energy efficient smartphone-based activity recognition using fixed-point arithmetic. J UCS 19:1295–1314
- Krishnan NC, Colbry D, Juillard C, Panchanathan S (2008) Real time human activity recognition using tri-axial accelerometers. Sensors, signals and information processing workshop. pp 3337–3340
Cited by
Sorted by citations
Schrader, L., Vargas Toro, A., Konietzny, S., Rüping, S., Schäpers, B., Steinböck, M., Krewer, C., Müller, F., Güttler, J., & Bock, T. (2020) · Journal of Population Ageing · doi:10.1007/s12062-020-09260-z · Cited by 110 · see citing works
- Authors
- Zhu, J., San-Segundo, R., & Pardo, J. M.
- Affiliations
- Beijing University of Posts and Telecommunications (CN)
- Universidad Politécnica de Madrid (ES)
- Corresponding author
- San-Segundo, Rubén
- Venue
- Human-centric Computing and Information Sciences
- Published
- 2017-06-02
- Volume
- 7
- Issue
- 1
- ISSN
- 2192-1962 (ISSN-L)
- OpenAlex
- W2608703604
- Semantic Scholar
- 5b179f0b69348dc6981fd06e975d9b51d52f87e7
- Cited by
- 68
- References
- 31
- Citations by year
- 2026: 1 · 2025: 6 · 2024: 11 · 2023: 5 · 2022: 12 · 2021: 8 · 2020: 8 · 2019: 10 · 2018: 7
- FWCI
- 2.84
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Locations
- Human-centric Computing and Information Sciences (publishedVersion, cc-by)
- UPM Digital Archive (Technical University of Madrid) (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- Funders
- China Scholarship Council
- Ministerio de Economía y Competitividad, Spain (TIN2014-54288-C4-1-R, DPI2014-53525-C3-2-R)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
- Physical Sciences › Engineering › Biomedical Engineering › Gait Recognition and Analysis
- Health Sciences › Medicine › Physiology › Physical Activity and Health
- Keywords
- human activity recognition, physical activity recognition, inertial measurement unit, wearable sensors, feature extraction, signal normalization, accelerometer, gyroscope, magnetometer, activity classification, Opportunity dataset
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
