Monitoring athletes with wearable sensors allows to gain insights on their technique and physical condition. However, invasive setups containing a large number of sensors may hinder the mobility of the athletes, leading to under-performance and possibly inaccurate data collection. In this paper, we show that correlation between data collected by diferent wearables can be used to identify a minimal setup. We propose a methodology to remove the least important sensors, and apply it to data collected by monitoring field hockey players. In this study, the number of sensors was reduced from 23 to 8 by deleting those exhibiting a correlation above 98%. Additionally, we demonstrate that even with a minimal sensor configuration, a significant amount of information is retained with regards to predicting the ball speed following a drag-flick, an important technique in ifeld hockey. Our experiments indicate that the utility of the data for this specific task remains practically unaltered.
The utilization of wearable sensors in the study of
athletes allows researchers to obtain a comprehensive
understanding of their physical condition. These sensors are
extensively employed to monitor various health-related
parameters, including workload, hydration levels, and
heart rate [
A typical setup usually involves the attachment of several IMUs to the subject using a specialized bodysuit. Despite the considerable value of these sensors as Published in the Workshop Proceedings of the EDBT/ICDT 2023 Joint Conference (March 28-March 31, 2023, Ioannina, Greece). ⋆You can use this document as the template for preparing your publication. We recommend using the latest version of the ceurart style. ∗Corresponding author. †These authors contributed equally. nEvelop-O (M. Gijssel) minimal setup. We also establish the groups of sensors that are greatly correlated and have limited value for data controllers. • Finally, we show that the obtained minimal sensor configuration can still efectively predict the ball speed during drag-flick, with only a slight reduction in performance as compared to the complete
Our experiments are conducted on a real-world sen- form of a player. It is typically measured by high-speed sor dataset, which was collected by field hockey players cameras or radar systems, and predicting speed based while practicing the drag-flick. on IMU data is an open problem. Rather than focusing on achieving the lowest possible prediction error, we use ball speed as a target variable to select important IMUs.
In this section we describe the relevant concepts of field hockey and stress the importance of feature selection from raw data.
Drag-flick . Our work focuses on data collected from ifeld hockey players during their practice of a technique known as the “drag-flick”.
The drag-flick is a highly efective shot in which the ball is lifted above the backboard of the goal rather than simply being hit, making the shot more aggressive at penalty corners. It is one of the most important techniques to be mastered in field hockey and requires coordinated action of multiple body parts. The key stages of a drag-flick can be seen in figure 1. After the acceleration phase T1-T3, T4 begins with the right foot rotating on the ground while the left foot leans as far forward as possible. At T5, the left foot lands on the ground. At the same time, the onset right wrist flexion begins with simultaneous maximal knee flexion, followed by extension. The follow-through of the shot occurs at T6 which concludes the drag-flick. Prior works have demonstrated that the greater the maximal knee flexion angle between T4 and T6, the higher the resulting ball speed. [9].
Examining sensor recordings of a drag-flick shot offers valuable insights into the player’s form. However, an overly invasive IMU setup can lead to subpar performance by the athlete, resulting in inaccurate data. As a result, drag-flick monitoring constitutes the perfect example of a task that could benefit from a reduced sensor setup.
Ball speed prediction. To measure the loss in utility of our minimal sensor setup compared to the original one, we evaluate the performance of ball speed prediction during drag-flick. The speed of the ball in field hockey is considered one of the most important metrics to assess the
In machine learning, feature selection is the task of identifying relevant input variables (features) in the data. Selecting important features improves the generalization capabilities of models and prevents overfitting. In the context of medical and sports research, however, feature selection can also be used to find relevant sensors from a larger pool of potential IMUs. This allows to reduce redundancy and increase the accuracy of the analysis by only focusing on the most informative IMUs.
Feature selection techniques are divided in three main types: filter methods, wrapper methods, and embedded methods [11]. Filter methods are “a priori”, meaning that they are performed independently of the chosen prediction model. These are usually based on metrics such as correlation or mutual information between individual features and the target variable. Wrapper and embedded methods, conversely, choose the most relevant features based on how much they contribute to the prediction for a specific model.
A comparison of these three types of techniques has been done by prior works [12, 13] and is beyond the scope of this paper. In our work, we aim to show that feature selection can be used to determine a minimal setup of wearable sensors. Adopting a minimal sensor setup has the two-fold advantage of reducing the equipment cost while making it less invasive for the athletes, leading to more accurate and realistic measurements. The feature selection algorithm that we adopted in our analysis is a combination of a univariate filter method and the correlation.
Body sensors and athlete monitoring. The advancement of technology and the continuous need for performance enhancement in sports [14] has accelerated the utilization of wearable sensors in sports data analytics [15, 16, 17]. A sensor network was introduced to monitor important indicators of world-class rowers for diferent rowing techniques [18]. Schmidt et al. utilized an IMU wearable setup to measure stance duration of both track and field sprinters [19]. The researches managed to identify four diferent types of basketball shot with 98% accuracy via a micro IMU-based wristband [20].
These devices have also been employed in field hockey, for example to recognize players’ activities [21] and to improve their technique. Iwamoto et al. [22] proposed a sensor-based approach to improve the push pass technique. Eight sensors were applied on several contact points of the stick and associated with diferent feedback sounds. This enabled players to receive real-time feedback on the efectiveness of their pass.
Furthermore, other works before ours focused on the study of the drag-flick technique. Body sensors were also utilized for assisting new players in executing a successful drag-flick [ 9]. The participants of the study have reported the improvement in the drag-flick technique, a decrease of the body load during the shot, and commended the more interactive learning experience. Minimal sensor setup and feature selection. Previous works have investigated reducing the number of devices in body sensor configurations. Typically, in these studies, relevant sensors were chosen through feature selection, resulting in an optimal setup for a particular task. In the field of digital healthcare, Caramia et al. [ 23] utilized feature selection to determine relevant sensors in the recognition of Parkinson’s disease using IMU-based devices. Their approach involved separating the sensors into various sub-groups and evaluating their significance in prediction. The significance was determined by calculating the accuracy of various classification algorithms. Khademi et al. [24] aimed to minimize the number of sensors in gait mode recognition. They proposed a novel gradient-based feature selection method, which introduces a penalty for the number of used features in the objective function. This method is applicable only to models that are trained based on gradient descent, such as neural networks. Amjad et al. [25] published a comparative evaluation of feature selection approaches for human activity recognition (HAR). At present, a minimal sensor setup can be realized via a single advanced compound device, such as a fitness tracker. A single wearable devices has been used in a number of lifelog studies, involving both regular participants [26] and professional athletes [27]. Utilizing only a single device for monitoring various parameters tends to be less invasive and equally informative for some applications.
Seniors
In our experiments, we utilize a private dataset collected by Kinetic Analysis1 in collaboration with the Royal Dutch Hockey Association (KNHB). The dataset comprises multiple trials per hockey player collected over four years (2017, 2018, 2019, and 2021), where each trial represents a full execution of a complete drag-flick. Since the dataset was collected over multiple years, henceforth we refer to it as DragEXT (drag-flick extended).
In total, 77 highly capable individuals of various ages participated in the data collection. The following age groups are represented in the dataset: U16 (younger than 16 years old), U18 (16-17 years old), U21 (18-20 years old), and seniors (older than 20 years old). The age and gender breakdown of the hockey players per each year is depicted in table 1.
Overall, more than 1, 500 drag-flicks have been recorded, with ball speed measured for at least 1, 300 trials. Around 350 drag-flicks have been manually annotated to contain timestamps for all the events typical of a drag-flick (as depicted in figure 1). The trials were recorded using the Moven suit by XSens. In total, 17 IMUs are contained in the suit to monitor various characteristics during the flick, including on the head, sternum, spine, and pelvis, as well as on shoulders, upper arms, forearms, hands, feet, and finally on upper and lower legs. These IMU measurements are combined to form a total of 23 sensor segments. Throughout the paper, we test our algorithm to reduce the number of such sensor segments, which we call just “sensors” for simplicity. We depict the sensors and their on-body placement in more details in table 2. In addition to objective measurements taken by accelerometers and gyroscopes, sensors were also calibrated to record derived measurements of position, velocity, orientation, and some other metrics. Measurements of position and velocity are recorded as 3D coordinates (x, y, and z), while the orientation is measured in quaternions2. Sensors were collecting data at the highest available sampling frequency of 240 Hz. Therefore, each trial in the dataset is represented as a set of 1https://www.kinetic-analysis.com/ 2https://base.xsens.com/s/article/MVNX-Version-4-File-Structure
In this section we describe our methodology for sensor selection, the feature extraction procedure and benchmarks for our models in great details.
Our sensor selection algorithm is divided into two main parts. In the first part, we extract features from sensor time series and we rank them individually, meaning that we only consider how much each feature is related to the target variable (in our case, the ball speed). Then, we construct a histogram with the top relevant features as in figure 3. This is used to do a preliminary ranking Table 2 of the sensors, based on the number of highly relevant The full list of sensors (segments) utilized in the Moven suit features that they produce. The extracted features and by XSens. The bottom half of the sensors are duplicated and the method used to rank them may vary depending on the placed on both parts of the body. In our reduction methodol- application. The specific method adopted in this study is ogy we aim to identify the most impactful. described in section 6.
In the second part, we discard redundant IMUs based on their correlation in the time domain. For example, by time series for all combinations of available sensors and examining the velocity measured by two sensors located recorded measurements, totalling 437. For the annotated on the spine (T8 and T12) in figure 4, one can easily see kicks it is feasible to associate the above time series with that they exhibit a strong correlation during the dragthe respective events of a drag-flick (T1-T6). lfick. When two sensors are highly correlated, we discard Preprocessing. As a preliminary step, we format the trial the one lower in ranking. This procedure aims at retainrecordings as time series data. For all combinations of ing the sensor which is more likely to produce relevant sensors, measurements, and coordinates in each trial, we features. extract a sequence of data and format them as a multidi- To eficiently select sensors for the minimal setup, mensional time series, as shown in figure 2. Throughout we first sort them by their ranking in terms of relevant the paper, we refer to the sequences forming the multidi- features, starting with the top ranked. We delete all mensional time series as “channels”. The name of each the lower-ranked sensors which are above a correlation channel follows the format “sensor_measurement_coor- threshold , which is a parameter of our algorithm. We dinate”, e.g., neck_velocity_x. then repeat the procedure for the non-deleted sensors, or
Furthermore, for the purpose of predicting the ball dered by their ranking. The remaining sensors at the end speed, we keep only the samples between the key frames have a pairwise correlation below the chosen threshold. T4 (right foot touch during pre-stretch phase) and T6 The complete procedure is detailed in algorithm 1. (end of the drag-flick). Indeed, this is a “greedy” algorithm, based on the intuition that sensors producing more relevant features provide higher utility. To find the truly optimal combination of sensors, one would need to try all the possible combinations of uncorrelated sensors, which however is Occurrences of top = 100 features
Algorithm 1 Sensor selection procedure SensorSelection(list of sensors ordered by ranking, threshold ) ←̄ for in do
▷ if not in ̄ then
continue for ′ in ̄ do return ̄ if ≠ ′ ∧ corr( , ′) > then
Remove ′ from ̄
Initialize the final list of sensors ▷ Skip to the next iteration 8 6 4 2 0 ) s / m ( y t i c o l
5 adnH eadH fteeoLT reegLpp rrepAm rreAm L g t h g i R eL rAm 3L 21T reegLw reegoLLw itegoThR trreFAm eaLdn cekN lisevP 8T lred ttooF t e
r rep re o ftH p ftoF t
tfoSuh ighR fteooFL loSduh U p p
U p ithR itghR fteLU g e e tfLU L tfoL th e L ig
R h g i
R measurements over trials, and 1̂, … , ̂ the corresponding predictions of a regression model, the MAE is computed as
MAE =
∑ | − ̂ | while the RMSE is computed as
RMSE =
∑( − ̂ )2.
Both metrics are important to provide an overall understanding of a regressor’s performance. The main difference between the two is that RMSE tends to be more sensitive than MAE to infrequent large errors, meaning that it can be significant if there are few outliers in the dataset. As a baseline for both metrics, we consider the performance of a naive regressor that always predicts the average ball speed (i.e., 81.52 km/h). This regressor yields a MAE of 9.68 km/h and a RMSE of 12.09 km/h. Any regression model providing higher error values should be considered non-informative.
In this section, we apply our sensor selection algorithm we utilize all the IMUs and keep the overall top-100 features. The precise procedure for feature extraction is detailed in section 6.1. (1) (2) 0 200 400 600
800
In order to test our sensor selection algorithm, we as- to the DragEXT dataset. To evaluate the full sensor setup, sess the performance of basic machine learning models that predict the ball speed before and after reducing the
L5
ead 3L 5L too r H tfF eL reo fteH e e L tfF m adn regL r e A
ld ftoT rm eg cek lisv too r L tfoLw ftoSu eL repA reLpp N eP itgFhR tregoFhRAm itagdnhRH itreeggoLLhRw litregoSduhhR itegoThR trreppAUm treegLppU 21T 8T e eL eLh tfeLpU fteLU i ighR ighR
In order to extract features from our time-series data, we utilize the TSFresh Python package3. With default parameters, TSFresh extracts 783 features for each channel of a multidimensional time series, including peak values, minima, autocorrelation, quantiles, and more. Counting all the sensors, measurements, and coordinates, each sible features were to be extracted, they would be over 340,000, making the dataset imbalanced in width and hard to process.
Therefore, as a preliminary selection, we keep only the velocity measurements. Since estimation of position, velocity, and acceleration over time provide an equal description of the motion, keeping only one of them should not reduce the amount of information in the time series. Regarding orientation, preliminary experiments showed that including it does not have a major impact on the results. Therefore, we decided to not include such measurements.
After extracting all the features from the velocity mea3https://tsfresh.readthedocs.io
Once the top 100 features are extracted, we evaluate the performance of multiple regression models using 10-fold cross-validation. Normally cross-validation is used to tune the hyperparameters of a model, but in this case we adopt it to evaluate models without choosing a specific train-test split. It is important to note that our goal is to assess the expected performance of several regressors,
Minimal setup. We ran our sensor selection algorithm using the absolute value of Pearson correlation for the correlation metric, i.e., for two sensors , ′ we computed corr(, ′) = |
∑=1 [], ′[] √(∑=1 []) 2(∑=1 ′[] 2) | ,
(3) where is the time index within the time series. The time series used for comparing sensors is the magnitude of the sensors’ velocity. In other applications, Pearson account for shifts or delays in possibly correlated time series. However, in the case of DragEXT, IMUs were taking trial’s timeseries comprises 437 channels. If all the pos- not to obtain a ready-to-deploy model. surements with TSFresh, we select the top = 100 rel- correlation may not be a suitable metric, since it does not
1.0 measurements synchronously, thus correlated time series prior works have considered linear models to study the were always aligned. The Pearson correlation computed relationship between lead-knee extension and ball speed between diferent sensors can be seen in figure 5. [10]. Finally, our algorithm produces a setup requiring
After eliminating all the sensors above a threshold one third of the sensors compared to the full set of IMUs, of 98% correlation, the minimal setup identified by our reducing them from 23 to 8. algorithm results in 8 sensors, namely those attached to left toe, right toe, left hand, right hand, left lower leg, right lower leg, left forearm, and T12 (located on 7. Discussion the spine). The correlation matrix between the IMUs in the minimal setup is shown in figure 6. Notably, the T8 Quality and quantity of sensors in the setup. Naturally, sensor has been removed, being highly correlated with the ideal minimal setup may difer depending on the task T12 as shown in figure 4. at hand. For example, some problems may require as rEergrorersssionrsbablelfsopreeeadnpdreadfteirctaipopnl.yWineg eovuarluseantesoar nseulmecbteiornof imnavnasyivveanrieosussomfIeMasUusr.emFoernotsthaesrpsoitssisibmleurcehgamrdolreessviotfalthtoe algorithms. These include widely-used machine learning ensure convenience and mobility of the athletes by wearmodels such as linear regression, decision trees, random ing limited number of the sensors. Since the proposed forests, and k-nearest neighbors. The complete list, along algorithm of sensor selection makes use of the correlawith the validation performance in terms of MAE and tion threshold , our approach enables data controllers RMSE, can be seen in table 3. to balance the tradeof between number of IMUs and the
The minimal setup obtained by our algorithm generally athletes’ comfort. provides a slightly higher error, both in terms of MAE and Varying sensor setups for diferent sports . Depending on RMSE, compared to the complete sensor setup. However, the required levels of mobility, the setup proposed for it is notable that such errors are still considerably smaller ifeld hockey may not be an ideal solution for other sports. than the baseline, indicating that most of the data utility For instance, for some sedentary disciplines like chess has been preserved. Additionally, some simple models and racing organizers tend to mandate athletes to wear such as ridge and lasso regression (which are essentially specific on-body sensors regardless of their invasiveregularized linear regression models) are outperforming ness [29, 30]. Therefore, the likely optimal setups are the complete setup. These models are typically more discipline-specific and need to be established separately. interpretable, providing practical insights. For instance, We believe our approach may be tuned to other sports
k-Nearest Neighbors
and encourage researchers from other domains to inves- are heavily correlated, retaining only the most important tigate the problem of sensor reduction. ones. We showed that it is feasible to reach comparaMirroring optimal sensors. Again, we establish the best ble results for ball speed prediction in field hockey with 8 sensors for the minimal setup to be those attached to both full and minimal sensor setups. We experimentally left/right toe, left/right hand, left/right lower leg, left demonstrated that for some of the models achieve better forearm, and T12 (spine-based). Interestingly, for those performance with the minimal setup. Our sensor selecsensors that do have counterparts on the opposite side tion approach can be employed for other sports and other of the body, our algorithm recognized both of them as tasks that require the use of IMUs. most important except for the one located on left forearm. Since not all the players who took part in the data collection were right-handed, it does appear logical for Acknowledgments the optimal setup to be mirrored with respect to both hands to not overemphasize the dominant hand for the This project has received funding from the European majority of the players. Therefore, we believe that de- Union’s Horizon 2020 research and innovation prospite the sensor on right forearm to not be selected by gramme under the Marie Skłodowska-Curie grant agreethe algorithm as the top-8 IMUs, it should be added to ment No 813162: RAIS – Real-time Analytics for the the final setup to ensure fairness with respect to minority Internet of Sports. The content of this paper reflects the players. views only of their author (s). The European CommisFuture work. The future research includes experimenting sion/ Research Executive Agency are not responsible for on other tasks in field hockey. Since, at present, the any use that may be made of the information it contains. problem of annotating the events during drag-flick is typically being done manually, we consider to extend References our approach to automatically detect stages of the shot.
We further plan to combine the research of minimizing the number of the sensors in the setup with the event segmentation tasks. Finally, we would like to collect a new dataset with both full and minimal sensors setups for the same users to assess the comfort of players and measure whether they exhibit any improvement during ifeld hockey tasks.
In this work we explored the possibility of reducing the required intertial measurement units (IMUs) to monitor field-hockey players. Specifically, we focused on a minimal setup to assess players’ form while executing a drag-flick. Our approach eliminates the sensors that [26] S. Yfantidou, C. Karagianni, S. Efstathiou, A. Vakali, J. Palotti, D. P. Giakatos, T. Marchioro, A. Kazlouski, E. Ferrari, Š. Girdzijauskas, Lifesnaps, a 4-month multi-modal dataset capturing unobtrusive snapshots of our lives in the wild, Scientific Data 9 (2022) 663. [27] V. Thambawita, S. A. Hicks, H. Borgli, H. K. Stensland, D. Jha, M. K. Svensen, S.-A. Pettersen, D. Johansen, H. D. Johansen, S. D. Pettersen, et al., Pmdata: a sports logging dataset, in: Proceedings of the 11th ACM Multimedia Systems Conference, 2020, pp. 231–236. [28] M. G. Kendall, The treatment of ties in ranking problems, Biometrika 33 (1945) 239–251. [29] FIDE, 2022 FIDE World Fischer Random Championship in Reykjavik, https://www.frchess.com/ regulations, 2022, October. Online; Retrieved January 30, 2023. [30] FIA, 2023 Formula 1 Technical Regulations, https://www.fia.com/sites/default/files/fia_2023_ formula_1_technical_regulations_-_issue_4_-_ 2022-12-07.pdf, 2022, December. Online; Retrieved January 30, 2023.