Human activity recognition (HAR) systems implement workflows that automatically detect activities from motion data, captured e.g. by wearable devices such as smartphones. These devices contain multiple sensors that record human motion as acceleration, rotation and orientation in long time series (TS) data. As a first step, HAR methods typically partition such recordings into smaller subsequences before applying feature extraction and classification. In this study, we evaluate the performance of 6 classical and recently published TS segmentation (TSS) algorithms on a new large HAR benchmark of 126 TS with up to 13 diferent activities, called MOSAD, recorded with 6 participants using ordinary smartphone sensors. Our results show that the ClaSP algorithm achieves significantly more accurate results compared to the other methods, scoring the best segmentations in 57 out of 126 TS. The FLOSS algorithm also shows promising results, particularly for long TS with many segments. MOSAD is freely available at https://github.com/ermshaua/mobile-sensing-human-activity-data-set.
Published in the Workshop Proceedings of the EDBT/ICDT 2023 Joint Conference (March 28-March 31, 2023, Ioannina, Greece). * Corresponding author. $ ermshaua@informatik.hu-berlin.de (A. Ermshaus); singhsun@informatik.hu-berlin.de (S. Singh); leser@informatik.hu-berlin.de (U. Leser) 0000-0002-8138-3060 (A. Ermshaus)
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License complexity of human behaviour. In contrast, MOSAD CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) includes long motion sequences that capture up to 13 diferent behaviours, which poses a challenging setting for TSS algorithms. Our specific contributions in this paper are: efect of the geomagnetic field on the device, providing insight into the orientation of activities. All three sensors produce triaxial measurements in the X, Y, and Z directions, and are typically digitized at a few hundred Hertz (Hz), resulting in long time series (TS) data that capture typical human activities (like standing or going) as homogenous segments. We formalize such recordings with the following concepts: 1. We present MOSAD, a human activity data set recorded with 3 mobile sensors from a smartphone. We recorded 20 diferent activities in 3 motion sequences with 6 participants, totalling 14 recordings of 9 triaxial sensor readings (126 TS), annotated and preprocessed the data, and make it Definition 1. A time series (TS) is a sequence of ∈ publicly available on our supporting website [16] N real values, = (1, . . . , ), ∈ R that contains the for follow-up works. observable output of a sensor over time. The values are also 2. In our experimental evaluation, we used MOSAD called observations or data points.
to assess 5 state-of-the-art TSS algorithms, TS are typically sampled at a fixed rate, such that the namely ClaSP [15], FLOSS [14], ESPRESSO [17], duration between two consecutive time points is always BinSeg [18] and BOCD [19] as well as a simple the same. This simplifies their inspection and is a prereqbaseline called Window [13]. ClaSP scores sig- uisite for many advanced analytics. The central property nificantly better results than the competitors and of a TS is that the measurement was recorded before achieves the most accurate segmentations in 57 +1 which leads to local patterns that may repeat, drift out of 126 cases. FLOSS scores the 2nd best seg- or suddenly change over time. This allows for the detecmentations and is very accurate for TS with many tion of underlying patterns, trends, and anomalies in the segments. data [23]. 3. We make a special efort in this work to make our used Python source codes and evaluation frame- Definition 2. Given a TS , a subsequence , of work, Jupyter-Notebooks, as well as all experi- with start ofset and end ofset consists of the contiguous ment data and visualizations publicly available observations of from position to position , i.e., , = on our supporting website [16] to foster the re- (, . . . , ) with 1 ≤ ≤ ≤ . The length of , is producibility of our findings and replicability for |,| = − + 1.
follow-up works.
Section 2 we introduce background knowledge and concepts used in this study, Section 3 discusses related works. In Section 4 we present MOSAD, Section 5 evaluates the TSS algorithms with it, and Section 6 summarizes our ifndings.
Definition 3. Given a TS that captures a motion
seHuman activity recognition (HAR) systems use methods quence, a change point (CP) is an ofset ∈ [1, . . . , ]
such as video analysis [20], environmental sensing [
The most commonly used and insightful wearable sen- as time series segmentation (TSS) and approached as an sors in smartphones for HAR are the accelerometer, gy- unsupervised learning problem [12]. In this work, we roscope, and magnetometer [22]. The accelerometer cap- study TSS in the context of human activity recognition tures the acceleration forces acting on a mobile device, (HAR), where it is used as a preliminary preprocessing which can indicate the presence or absence of motion. step before feature extraction and classification. The goal The gyroscope measures the impact of gravity on the of TSS in this context is to partition the sensor recording device as angular velocity, allowing for estimation of ro- into consecutive segments, which can then be further tation during movement. The magnetometer records the processed and used to identify individual activities.
ably, and also refer to their length as the width. Periodic TS repeat similar subsequences of a fixed length, which we call temporal patterns (or periods). However, local parts of TS may still deviate from each other, such as in period length, shape or amplitude, and temporal patterns can drift or change over time. This characteristic of TS makes segmentation challenging, as it requires the identification of patterns that may vary over time.
Human activity recognition (HAR) has been widely
studied over the past few decades and has seen many
methodical improvements [
The literature contains many specific HAR workflows
that combine feature engineering and classification
techniques [
The TSS literature contains many methods applicable focus on large number of segments. We find this espeto activity segmentation, as surveyed in [12]. Bayesian cially important for HAR, as human behaviour is complex, approaches split TS into windows and compare their evolving and the technology used to capture it frequently probability distributions to infer CPs. A popular imple- changes. mentation is Bayesian Online Change Point Detection (BOCD) [19], that uses a recursive message-passing algorithm to infer the most recent CP. It has also been ex- 4. MOSAD tended for short, gradual changes [29]. Another branch of TSS solves optimization problems to induce segmen- We introduce MOSAD (Mobile Sensing Human Activity tations. Given a user-selected cost function that defines Data Set), a new multi-modal, annotated TS data set that the notion of segment homogeneity and a search func- contains 14 recordings of 9 triaxial smartphone sensor tion, the problem can be numerically solved. A catalogue measurements (126 TS) from 6 human subjects performof parametric and non-parametric cost functions have ing (in part) 3 motion sequences in diferent locations. been proposed [13], and ensembled to increase model The aim of the data set is to facilitate the study of human robustness and accuracy [30]. Accurate exact and approx- behaviour and the design of TS data mining technology imate search functions include Pruned Exact Linear Time to separate individual activities using low-cost sensors in (PELT) [31] and Binary Segmentation (BinSeg) [18]. wearable devices. In creating the data set, we focused on
Recently, TSS methods that impose no assumptions on capturing pervasive motion sequences to record meaningthe observed data points and change types have been pub- ful human behaviour. For data collection, we utilized the lished and evaluated on large benchmarks [15]. FLOSS built-in sensors of a smartphone to record the subjects in measures the density of similar subsequences in poten- an unobtrusive and realistic setting, as is commonly done tial segments and greedily extracts the requested amount by many people. We annotated the transitions between of CPs [14]. ESPRESSO extends FLOSS with TS chains, activities in the recordings and sampled successive data positional subsequence information and a more sophisti- points at a fixed rate of 50 Hz to enable machine learncated entropy-based segmentation procedure [17]. Lastly, ing technology to analyse the measurements and extract ClaSP is one of the most recently published TSS algo- knowledge. rithms that formulizes TSS as a collection of hypothetical, In the following subsections, we elaborate in detail self-supervised TS classification (TSC) problems, where on the data set design (Section 4.1), the data collection the best-performing label configuration induces the seg- process (Subsection 4.2), and the annotation and pre
Time Series Sensor Data
Annotation
Removing Pauses
Sampling
Motion Sequence
1 1,2,3 1,2,3 1,2,3 1,2,3
1 processing steps (Subsection 4.3). We will conclude our We used the application "Physics Toolbox Sensor Suit" technical description of the data set with an overview to record the phone’s accelerometer and gyroscope data of its specifications and examples (Subsection 4.4) and from the IMU component (type "LSM6DSL"), as well as its information on its availability (Subsection 4.5). magnetometer (type "YAS539"). Additionally, we filmed each subject with another smartphone to capture the ground truth movements. The data set includes mea4.1. Data Set Design surements from all three (X, Y, and Z) sensor axes, and The primary objective of MOSAD is to capture the nat- is annotated with activity transitions, resulting in a 9ural dynamics of human behaviour. To achieve this, we dimensional multivariate TS and one list of CP positions used activities that are widely practised and arranged per subject and routine. them in an order based on the places where they were recorded. Human behaviour is not entirely unstructured, 4.2. Data Collection as we tend to group similar activities and repeat them periodically. Therefore, recording a common pre-arranged routine is a sensible choice, and it has the advantage of being comparable across subjects and modalities. We designed three motion sequences, comprising 13, 3 and 6 activities, respectively. Two of the sequences were recorded indoors, while the third was recorded outdoors.
Figure 1 illustrates pictures of some of the items and places used. The routines consist of a total of 20 diferent activities, varying in length from a few seconds to a few minutes. The following enumeration lists the ordered activities: Before the recordings, we provided instructions on how to perform the motion sequences and set time constraints for each activity. We began the data collection with a preparation phase, in which the smartphone was placed in the subject’s pocket, filmed the routine, and ended with a follow-up phase, where the mobile device was eventually removed. During the recording, we gave auditory cues to the subjects to assist with timing and maintaining the order of activities. The subjects were asked to pause briefly between two consecutive activities to prevent transitional movements. These pauses were later removed as part of the preprocessing. 1. Household (indoor): descend stairs, climb stairs, We encountered multiple problems throughout the vacuum, lie, iron, mop, sit, make bed, stand, slow data collection process, that we want to share to help walk, hang out laundry, walk, fold laundry researchers with follow-up works. In the initial experi2. Sport (indoor): spin, sit-ups, modified push-ups mentation phase, we used an iPhone 11 Pro, which ran3. Sport (outdoor): slow walk, run, walk, rope jump, domly crashed the application during the recordings, squat, jumping jack resulting in lost measurements. To overcome this issue, we switched to an Android smartphone, which also in
These routines were performed, in part, by 6 human creased the sensor sampling rate from 100 to 300 Hz.
subjects. See Table 1 for more information. The partic- Additionally, the automatic standby mode of the
smartipants comprise 2 males and 4 females, which cluster phone resulted in data loss. This can be addressed by
into young and mid-aged adults. Such an age gap can be using other applications that prevent the smartphone
interesting to study, as movement changes with age [
All participants were recorded using a Samsung Galaxy M20, which was placed in their front right trouser pocket. In the preprocessing stage, we carried out three key transThis location is common for smartphones and has been formations to create an annotated and consistent data set. shown to be highly accurate for activity recognition [32]. Figure 2 illustrates this workflow. Firstly, we annotated the sensor signals with timestamps at which activities change to obtain a ground truth. We use these CPs in the evaluation to assess the quality of the TSS methods.
To create the annotations, we carefully examined the
videos in conjunction with the sensor measurements to
derive precise ofsets that capture the exact time points
of activity transitions. Secondly, we removed the data
points from the preparation phase, pauses between
activities and the follow-up phase to clean the signals of
unwanted noise. Finally, we downsampled the
measurements, using linear interpolation, to a fixed sample rate
of 50 Hz. This is a common preprocessing step for sensor
signals, as their sampling rate often varies due to
physical component inaccuracies. It is necessary to create
a multivariate TS that is synchronized throughout time
and dimensions, improves interpretability as well as the
performance of TS analytics that assume equidistant time
gaps between measurements. We selected a sample rate
of 50 Hz, as used in the mHealth data set [
Activity Household Sport (indoor)
Sport (outdoor) 0.0m 3.3m 6.7m 10.0m 13.3m 16.7m
20.0m 4.4. Data Set Overview our setup in Subsection 5.1, then compare the accuracy of 6 TSS competitors in Subsection 5.2. To further demonIn total, MOSAD comprises 14 9-dimensional sensor strate the characteristics of our data set, we discuss the recordings, totalling 126 TS, from the 6 participants per- results of the two best-performing methods using a seforming up to 3 motion sequences. On average, the rou- lected example in Subsection 5.3. We make all used source tines 1-3 contain 19.9, 3.9 and 6.5 minutes worth of ac- codes, Jupyter-Notebooks, and the raw measurement tivity data, with only small deviations per participant sheets available on our website [16]. (Figure 3 top left). The single activities last between 12 seconds for descending the stairs and 203 seconds of vacuuming, also with small variances except for (slow) 5.1. Benchmark Setup walking (Figure 3 top right), the only two activities per- We conducted all experiments on an Intel Xeon E7-4830 formed in two routines (household and outdoor sport), with a clock speed of 2.20 GHz, 440 GB of RAM, and 80 yet with diferent time constraints. cores, using Python 3.8.
In Figure 3 bottom, we illustrate, as an example, the Xaxis acceleration recordings of participant 2 performing the 3 motion sequences. The single activities are coloured and show subtle but also more pronounced diferences between each other.
SA licence. Users are permitted to share and adapt the data with author credit, but must also use this licence. The data set can be downloaded on our supporting website [16], including a data loader for Python and supplementary materials.
of diferent TSS methods on MOSAD. We first describe Competitors We compare 6 established and recently published TSS algorithms on MOSAD. We include a simple baseline called Window [13] that measures the discrepancy between successive subsequences using a Mahalanobis cost function. As the width of the subsequences, we chose 5 seconds of motion data (250 data points), which is suficient to capture the transitions of human behaviour. We also include two classical methods, BinSeg [18] with auto-regressive cost (optimization-based) and BOCD [19] (Bayesian), that are very popular and have performed well in a recent evaluation [33]. Additionally, we include the density-based algorithm FLOSS [14] and its variant ESPRESSO [17]. Both require a window size, which should roughly capture one instance of a temporal pattern, that we set to 1 second (50 data points).
FLOSS also uses a sliding window to achieve more accurate results in long TS data that we set to 20 seconds of sensor data (1k data points). ESPRESSO needs a TS chain length that we set to 3 as in [15]. Lastly, we report results for our own method ClaSP [15]. Except BOCD and ESPRESSO, all the aforementioned algorithms use a minimum segment size parameter that we set to 5 seconds. This increases their accuracy, but it exhibits domain knowledge. We also set the number of segments as a hyper-parameter for all algorithms, as only ESPRESSO and ClaSP can automatically infer this information without adaptation. BOCD Window BinSeg
ClaSP FELSOPSRSESSO
Evaluation Metric The quantitative evaluation of TSS mean median std algorithms is challenging as signals and ground truth ClaSP 74.0% 75.0% 21.1% annotations can be ambiguous. Therefore, the literature FLOSS 66.3% 64.0% 14.4% contains a range of evaluation measures that quantify ESPRESSO 61.8% 60.3% 15.8% diferent notions of quality. These measures can be di- BinSeg 62.9% 60.6% 23.5% vided into classification-based and clustering-based ap- Window 56.1% 54.4% 18.6% proaches. BOCD 50.2% 49.2% 16.0%
While classification-based approaches check if a de- Table 2 tected CP matches the annotation (with some slack), Summary Covering performances for the 6 competitors on clustering-based approaches report the exact degree of the 126 TS in MOSAD. ClaSP shows by far the best results. deviation. In this study, we have chosen to use the latter and have selected the Covering measure [33]. This measure quantifies the overlap between predicted and 5.2. Segmentation Performance annotated segments. It is defined as follows:
Let the interval of two successive CPs [ , . . . , +1 ] denote a segment in and let as well as be the sets of predicted or ground truth segmentations, respectively. For notational convenience, we always consider 0 = 0 as the first and = + 1 as the last CP to include the first (last) segment. The Covering score reports the best-scoring weighted overlap between a ground truth and a predicted segmentation (using the Jaccard index) as a normed value in the interval [0, . . . , 1] with higher being better (equation 1).
tors on the entire MOSAD data set. The CD diagram in Figure 4 (left) shows that ClaSP (2.2) ranks first, followed by FLOSS (2.9), ESPRESSO (3.4), BinSeg (3.5), Window (4.1), and BOCD (4.9). FLOSS, ESPRESSO and BinSeg form a group of insignificantly diferent approaches, while Window and BOCD perform much worse. Considering the wins or ties (first position in ranking), ClaSP scores the best performances for 57 TS, followed by FLOSS (36), ESPRESSO (17), Window (10), BinSeg (8), and BOCD without any wins (counts do not sum up to 126 due to = 1 ∑︁ max ‖ ∩ ′‖ tCielasS).P Wraenkasn2anlydsepdostihtieonproorpwerotrisees, obfutthdeid6n9oTtSfindwahneyre ‖ ‖ ∈ ‖‖ · ′∈ ‖ ∪ ′‖ obvious commonalities, such as certain routines, sensors, (1) or subjects.
It is defined for sets with varying sizes (including being Considering these dimensions in isolation, we observe empty). In order to perform a performance comparison, similar results compared to the global ranking. For the we run the TSS algorithms with each of the 126 TS and single routines, FLOSS scores the best results for the aggregate the resulting Covering scores into a single household sequence, while ClaSP wins both the indoor ranking. First, we compute the rank of the score of each and outdoor sport motions. This indicates that the tempomethod per TS, where the best method is assigned rank 1, ral constraint in FLOSS, if properly set, enables it to score the 2nd best method is assigned rank 2, and so on. Then, good results in long activity recordings. We also analwe average the ranks of a method on all TS to obtain ysed the 42 TS per sensor type individually and find that its overall rank. To visualize the final ranking, we use ClaSP scores best-ranking results for all three of them, critical diference (CD) diagrams, as introduced in [ 34]. namely acceleration, gyroscope and magnetometer. A The best-ranking approaches with the lowest (average) similar finding holds if we aggregate our results by subranks are shown to the right of the diagram; see, for ject; ClaSP scores the best segmentation performances instance, Figure 4 (left). Groups of approaches that are for participant 1-5 while FLOSS wins for subject 6 (who not significantly diferent in their ranks are connected only performed the household routine). The rankings by a bar, based on a Nemenyi two-tailed significance test show that ClaSP scores state-of-the-art performances with = 0.05. across TS and their characteristics. FLOSS also scores good results and outperforms ClaSP in certain specific settings, though without a significant diference.
The summary statistics in Table 2 and Figure 4 (right) validate our findings. ClaSP scores the highest mean Covering performance of 74.0% with a rather large standard deviation of 21.1%. This can be explained by a large performance diference of circa 30 percentage points (pp) between the household and sport routines. On average, ClaSP improves compared to the 2nd best competitor FLOSS by 7.7 pp. In a pairwise comparison of ClaSP against the other algorithms, it achieves between 79 wins against FLOSS and 110 wins vs BOCD. However, its average performance still is only 74.0%, which needs improvement.
ity of computation for the best-ranking methods, ClaSP and FLOSS, by human inspection. Figure 5 shows a selected example of participant 2 performing the outdoor sport routine 3 (top) and the score profiles from both algorithms, including CP predictions (green vertical lines). ClaSP assigns high scores to TS ofsets that are likely CPs, while FLOSS annotates low scores. Local maxima (minima) are then extracted with diferent peak detection algorithms to locate the CP predictions. Both algorithms produce good segmentations in this example. ClaSP detects all 5 activity transitions, while FLOSS locates 3. The profiles, however, are quite diferent. ClaSP shows smooth scores with only small deflections, while FLOSS has many smaller peaks and is noisy. This complicates the CP detection and leads to worse results.
In this paper, we have introduced a new mobile sensing human activity data set, called MOSAD, which includes 126 TS from 6 participants performing three motion sequences. The data set utilizes current sensor technology and captures two age groups. It is freely available on our supporting website [16] and can be used in follow-up research.
We have also evaluated 6 state-of-the-art TSS algorithms using MOSAD, and find that the ClaSP algorithm significantly outperforms its competitors. These findings are consistent with recently published benchmark results [15]. It achieves the highest accuracy in 57 out of the 126 TSS in MOSAD, increasing the average accuracy by 7.7 pp compared to the 2nd best method.
Based on our results, we conclude that the ClaSP algorithm can be a suitable method for activity segmentation.
However, we acknowledge that there is still room for improvement in performance. Future work will investigate the use of multivariate TSS methods to exploit temporal dependencies between sensors. novel open framework for agile development of mo- logical and environmental sensors: A comprehenbile health applications, BioMedical Engineering sive survey, IEEE Access 8 (2020) 210816–210836.
OnLine 14 (2015) S6 – S6. [23] A. Ermshaus, P. Schäfer, U. Leser, Window size [10] A. Reiss, D. Stricker, Introducing a new bench- selection in unsupervised time series analytics: A marked dataset for activity monitoring, 2012 16th review and benchmark, 7th Workshop on Advanced International Symposium on Wearable Computers Analytics and Learning on Temporal Data (2022). (2012) 108–109. [24] C. R. Wren, E. M. Tapia, Toward scalable activity [11] D. Roggen, A. Calatroni, M. Rossi, T. Holleczek, recognition for sensor networks, in: Location- and K. Förster, G. Tröster, P. Lukowicz, D. Bannach, Context-Awareness, 2006.
G. Pirkl, A. Ferscha, J. Doppler, C. Holzmann, [25] A. M. Swartz, S. Strath, D. Bassett, W. L. O’Brien, M. Kurz, G. Holl, R. Chavarriaga, H. Sagha, H. Bay- G. A. King, B. E. Ainsworth, Estimation of energy ati, M. Creatura, J. del R. Millán, Collecting complex expenditure using csa accelerometers at hip and activity datasets in highly rich networked sensor wrist sites., Medicine and science in sports and environments, 2010 Seventh International Confer- exercise 32 9 Suppl (2000) S450–6. ence on Networked Sensing Systems (INSS) (2010) [26] N. Kawaguchi, Y. G. Yang, T. Yang, N. Ogawa, 233–240. Y. Iwasaki, K. Kaji, T. Terada, K. Murao, S. Inoue, [12] S. Aminikhanghahi, D. J. Cook, A survey of meth- Y. Kawahara, Y. Sumi, N. Nishio, Hasc2011corpus: ods for time series change point detection, Knowl- towards the common ground of human activity edge and Information Systems 51 (2017) 339–367. recognition, in: Ubiquitous Computing, 2011. [13] C. Truong, L. Oudre, N. Vayatis, Selective review [27] R. Chavarriaga, H. Sagha, A. Calatroni, S. T. Diguof ofline change point detection methods, Signal marti, G. Tröster, J. del R. Millán, D. Roggen, The Processing 167 (2020) 107299. opportunity challenge: A benchmark database for [14] S. Gharghabi, C.-C. M. Yeh, Y. Ding, W. Ding, P. R. on-body sensor-based activity recognition, Pattern Hibbing, S. R. LaMunion, A. Kaplan, S. E. Crouter, Recognit. Lett. 34 (2013) 2033–2042. E. J. Keogh, Domain agnostic online semantic seg- [28] A. Khan, Y.-K. Lee, S. Lee, T.-S. Kim, A triaxial mentation for multi-dimensional time series, Data accelerometer-based physical-activity recognition Mining and Knowledge Discovery 33 (2018) 96 – via augmented-signal features and a hierarchical 130. recognizer, IEEE Transactions on Information Tech[15] A. Ermshaus, P. Schäfer, U. Leser, Clasp: parameter- nology in Biomedicine 14 (2010) 1166–1172. free time series segmentation, Data Mining and [29] E. Draayer, H. Cao, Y. Hao, Reevaluating the Knowledge Discovery (2023). change point detection problem with segment[16] MOSAD, Code and Raw Re- based bayesian online detection, Proceedings of the sults, https://github.com/ermshaua/ 30th ACM International Conference on Information mobile-sensing-human-activity-data-set, 2023. & Knowledge Management (2021). [17] S. Deldari, D. V. Smith, A. Sadri, F. D. Salim, [30] I. D. Katser, V. Kozitsin, V. Lobachev, I. Maksimov, Espresso: Entropy and shape aware time-series Unsupervised ofline changepoint detection ensemsegmentation for processing heterogeneous sensor bles, Applied Sciences 11 (2021) 4280. data, Proc. ACM Interact. Mob. Wearable Ubiqui- [31] R. Killick, P. Fearnhead, I. A. Eckley, Optimal detectous Technol. 4 (2020) 77:1–77:24. tion of changepoints with a linear computational [18] A. K. Sen, M. S. Srivastava, On tests for detecting cost, Journal of the American Statistical Association change in mean, Annals of Statistics 3 (1975) 98– 107 (2012) 1590 – 1598.
108. [32] G. Bieber, J. Voskamp, B. Urban, Activity recogni[19] R. P. Adams, D. J. MacKay, Bayesian online change- tion for everyday life on mobile phones, in: Interpoint detection, arXiv preprint arXiv:0710.3742 acción, 2009.
(2007). [33] G. J. J. van den Burg, C. K. I. Williams, An evalua[20] M. A. R. Ahad, Vision and sensor-based human ac- tion of change point detection algorithms, ArXiv tivity recognition: challenges ahead, in: Advance- abs/2003.06222 (2020). ments in Instrumentation and Control in Applied [34] J. Demšar, Statistical Comparisons of Classifiers System Applications, IGI Global, 2020, pp. 17–35. over Multiple Data Sets, The Journal of Machine [21] S. A. Elkader, M. Barlow, E. Lakshika, Wearable sen- Learning Research 7 (2006) 1–30. sors for recognizing individuals undertaking daily activities, Proceedings of the 2018 ACM International Symposium on Wearable Computers (2018). [22] F. Demrozi, G. Pravadelli, A. Bihorac, P. Rashidi,
Human activity recognition using inertial, physio