Human activity recognition (HAR) is an active field of research due to the emerging applications in areas such as ambient assisted living (AAL), rehabilitation monitoring, fall detection, remote control of machines/games and analysing fitness data. The classification of human activities is frequently the key issue to be tackled. This is an interesting and challenging task, as activities from different subjects in different environments have to be recognized as the same class. A typical HAR pipeline consists of several steps: pre-processing, feature extraction, dimensionality reduction and classification.

The aim of our work is to develop a HAR system carried by masseurs for controlling a therapy table via different movements of legs or hip. For the masseur, it is important that no hands are involved, as they are mostly oily, and it is more comfortable if the therapy table can be operated remotely without stationary foot pedals. A voice control is also not an optimal choice, as the patients lying on the therapy table should be able to relax without disturbance. In our experiments, we studied two different sensor positions: fixed at the right hip like a belt and loosely inserted in a pocket of the trousers.

This study aims to be an advisor for creating classification models for HAR systems with wearable sensors and an additional help for finding optimal parameter settings. In Fig. , we depicted some influencing variables on the classification accuracy of a HAR system (see enumerations at the figure bottom). We split the influences according to their occurrence in the classification process. Most of these influencing variables have been studied in this work such as optimal window lengths and shifts, good features and how to select them, what is the minimal size for the training database, is a preceding teaching process necessary, and so forth. Figure  also shows that, due to the huge number of influencing variables, it is not easy to fully understand a classification process, and it is only related with much work. Therefore, the comparability between different classification tasks also becomes difficult.

The rest of the paper is organized as follows. Section  summarizes relevant work on human activity recognition. Section  explains the hardware and software infrastructure used in our HAR system and declares the used data collection processes. In the Sects. –, results of our experiments with the three different hardware systems are presented and discussed. In Sect. , a short comparison of the best-reached classification accuracies with the different hardware systems is given. Finally, Sect.  concludes the paper.

In recent literature, smartphones often serve as a tool for implementing a HAR system , as they are equipped with a rich set of sensors. However, device independence to remedy varying hardware configurations as well as efficient classifiers to prolong battery life and limit memory usage are still problems to be tackled.

Other challenges in the field of activity recognition are the differences in the way people perform activities concerning speed and accuracy (user independence) as well as the sensor positioning on the human body to find a position with high information gain and good separable features. Orientation independence of the sensor is also often desirable.

In , the difficulty of comparing HAR accuracy throughout literature is mentioned, since implementations vary across subject health or functional impairment, number of sensors and their placement locations on the body, activities performed, number and type of classes to distinguish, and validation techniques used. Additionally, various sensors may record with different measurement accuracies. For validating results of trained models, some papers use an n-fold process where samples from all subjects are blended. An alternative and better scheme involves the leave-one-subject-out (LOSO) technique so that the test set contains data of unseen subjects. In , classification accuracy drops from 100 % to 78.35 % when using the LOSO technique instead of splitting training and test data into 70 % : 30 % or using 10-fold cross-validation.

The optimal sensor placement on the human body plays an important role. Comparisons of different sensor positions have been considered in , , and . In the ankle was the best sensor position, while in and the trouser side pocket and the right thigh, respectively, were the best choices. However, in the shirt pocket was better than the right trouser front pocket and two further sensor positions.

In our work, we have chosen the trouser pockets as a desirable unrestricted sensor position. In literature, we found some publications that also used this sensor placement , in addition to multiple publications with other sensor positions. In Table , these publications are listed with additional information about type and placement of sensors, number of activities, used classification methods and their performance.

In Table , the following abbreviations are used: acc (acceleration), gyr (gyroscope), mag (magnetometer), bar (barometer), app (approximately), pos (position), CNN (convolutional neural network), KNN (K nearest neighbour), SVM (support vector machine), NB (naïve Bayes), DT (decision tree), RF (random forest), J48 (decision tree – J48), BN (Bayesian network), BDM (Bayesian decision-making), RBA (rule-based algorithm), LSM (least-squares method), DTW (dynamic time warping), ANN (artificial neural network), LR (logistic regression), IBK (instance-based classifier with parameter K; same as KNN), MLP (multi-layer perceptron), DTa (decision table), PNN (probabilistic neural network), ERMS (root-mean-square error).

Some publications listed in this table fixed the sensor at the thigh , and three further publications examined a sensor placement on the wrist similar to a watch and in the hand, respectively, . Other sensor positions like ankle can be found in the column sensor positions. If the authors used smartphones for the data collection process, we added the word “smartphone” in the “Sensor types” column in brackets.

If we take a closer look at Table , it is apparent that the usage of a single sensor type (accelerometer) achieves good classification performance. In , the additional use of gyroscope and magnetometer data resulted in an improvement in performance of 1.2 %.

A further interesting fact of Table  is that the number of activities to be classified do not negatively correlate with the classification performance. Also, publications with a high number of daily living activities as and gained comparable results in performance.

In literature, common classification methods are support vector machines as well as k-nearest-neighbour (KNN) approaches. In and , SVMs achieved the best results; on the other hand, in and , KNN methods performed best. So, there is no overall method that performs best regardless of circumstances. In , , and , the idea to use several stages for classification instead of a single model approach turned out to be advantageous. In , for instance, a two-phase approach improved the overall system by 9 % on average, whereas a combination of several classifiers led to an improvement of 7 %. In , the use of linear acceleration (excluding the effect of gravitational force) combined with the conversion of accelerometer readings from a body coordinate system to earth coordinate system gave huge improvements in classification accuracy.

In the following, we want to summarize – for the particularly interested reader – the uniqueness, strengths and weaknesses for every paper listed in Table .

described a wearable sensor system that is fixed above and below both knees. Data were logged at 25 Hz, which was sufficient to measure the lower extremities. A special approach to solve the problem of classification is the quaternion orientation representation of a body's orientation in three dimensional space. This information is transformed into colour images that serve as input for CNNs to get an automatic feature extraction. The LOSO validation technique was used. During data processing, it was found out that the body-worn sensors had slipped substantially on the legs of two subjects during cycling activities and on one subject during running activities. These trials had to be excluded from further evaluations. In this study, static tasks were better classified as dynamic tasks. The classification of data from subjects during cycling, ascending and descending stairs turned out to be difficult. Moreover, the different execution speeds of the subjects made it difficult to classify walking and running samples. Furthermore, the authors stated that battery lifetime would be better for single sensor systems using only accelerometers.

proposed a system without the need for pre-processing, as they use an end-to-end solution via CNN only. They used a publicly available dataset and three validation techniques, also regarding the robust LOSO validation. They stated that the minimum number of acceleration sensors required for perfect classification accuracy is two, where at least one of the sensors should be located on a body part with a wide range of motion such as the ankle.

made tests with lightweight classifiers (decision tree, KNN) while minimizing resource use. The KNN method, combined with data of accelerometer–gyroscope and a subset of eight features (found via the “ReliefF” feature selection algorithm), was sufficient for recognition. Frequency domain features were not necessary, as classifications with solely time domain features gave the best results. Accuracy also reduced slightly when a magnetometer was added to the accelerometer–gyroscope combination. For distinguishing upstairs and downstairs activities, the skewness feature of the accelerometer data was useful. For future analysis, additional physical and statistical features as well as other feature selection techniques would be interesting.

developed a HAR system that should perform robustly in any position that a smartphone could be held and which should be suitable for real-time requirements (sampling window size of 3 s). This goal was reached with an accuracy of more than 82 % for each activity in each of the four possible holding positions. The authors wrote that wearing multiple sensors on different body parts would be uncomfortable. The best results were achieved by SVM classifiers combined with the trouser side pocket. Furthermore, new features are extracted from a tri-axis accelerometer and a barometer like the variance of atmospheric pressure, and afterwards the features got normalized to a standard range of -1 to 1. So, they found features that are capable of distinguishing walking, ascending and descending stairs better than features of conventional methods. Moreover, the tracking of the subjects in a building completed the study. Nevertheless, analyses were only made with 10-fold cross-validation and with all 118 features.

implemented a mobile application to report the daily exercises of a user to estimate calorie consumption. The system is energy-efficient, runs offline and has only 10 kB training data to be stored. The experiments showed that 10 instances of each activity class are enough to get the same accuracies with the best classifier (KNN method, K = 1). Six classifiers and two feature selection algorithms showed that 15 out of 70 normalized, time domain features are enough to get the best success rate. As the system had slight difficulties to classify ascending and descending stairs, the authors propose to combine these two classes, if possible, into one class. Points to be improved in future are the fixed location and orientation of the phone in the trouser pocket, the use of a 10 s windowing process and the sole use of the 10-fold cross-validation method.

made a comprehensive study that compares eight classification methods in terms of correct differentiation rates, confusion matrices, computational costs as well as training and storing requirements. The classifier BDM (Bayesian decision-making) achieved the highest classification rates with RRSS (repeated random sub-sampling) and 10-fold cross-validation, whereas the LOSO technique recommended SVM, followed by KNN (with K = 7). In this study, 19 activities were classified via acceleration, gyroscope and magnetometer readings of sensors placed on five different body places of eight young subjects. Sampling frequency was set to 25 Hz, and data were segmented into 5 s windows. A big feature set was reduced via principal component analysis from 1170 normalized features to 30. The utilized receiver operating characteristic (ROC) curves presented, in a very clear way, which activities are mostly confused. Also, the analysis made with reduced sensor sets as well as the long list of potential application areas of HAR systems is very interesting. Moreover, for future investigations, the authors propose the development of normalization between the way different individuals perform the same activities. New techniques need to be developed that involve time wrapping and projections of signals as well as comparing their differentials.

specifically focused on the challenges of user-, device- and orientation-independent activity recognition on mobile phones. The orientation tests (same phone in vertical/horizontal orientation) especially showed that conventional HAR systems using phone coordinate systems are not robust enough. Firstly, proposed to use time domain, frequency domain and autocorrelation-based features. Secondly, linear acceleration that excluded the effects of gravitational force increased the accuracy further. Thirdly, the switch to Earth coordinates further boosted the classification accuracies up to 97 %. The only drawback of this method is the higher energy consumption for using three sensors, which will be further investigated by the authors.

used simple time domain features, such as the mean of the logarithm of each acceleration axis, to implement a two-phase activity recognition system which is device independent as well as position independent. Therefore, six different smartphones and four different positions of the devices are used. It is shown that the pattern for one activity varies from one device to the other, which makes, for example, threshold-based approaches unsuitable. Noise and outliers were eliminated, and filtering was conducted to preserve medium-frequency signal components. Phase 1 of activity classification chooses the best training dataset that yields maximum overall accuracy for a test set. Phase 2 further improves the accuracy by condition-based parameter tuning of a given classifier. As shirt pocket was found to be the best position for collecting training data, training data were only collected with this specific position.

tried to distinguish fall and non-fall events by computing four time domain features from accelerometer data and by using a two-layer feed-forward network. Data are collected at two different carrying positions of the smartphone; 26 fall and 41 non-fall events were classified with 93 % accuracy. In future, this study should be further extended to more subjects (now 2), to a bigger feature database and to more classifiers.

used a wearable device in a watch style for classifying five daily activities with an acceptable accuracy. A big advantage is the comfortable way of wearing the sensor. Unfortunately, sitting or bicycling activities cannot be detected reliably. For pre-processing, a median filter for eliminating noise and a temperature sensor for correcting the acceleration data were used. Afterwards, several time domain and frequency domain features were computed, and a decision tree was applied. To be more energy efficient, the micro-controller was set to a sleeping mode and only woke up when necessary. Weaknesses of the analyses are that only young people participated in the study, and the classification decision tree was only created from data of all persons together. Furthermore, analyses including feature selection methods or other classifiers would be interesting.

applied a divide-and-conquer strategy for first differentiating dynamic activities from static activities (threshold-based), and then posture recognition was used to classify the static activities sitting and standing. Therefore, thresholds for pitch and roll angles of the ankle were used. For dynamic activities, the exercise classification algorithm classified them into three classes: running, cycling and ambulation activities. Afterwards, the ambulation classification algorithm was utilized to distinguish between level walking, walking upstairs and walking downstairs. Here, two KNN classifiers with K = 5 were used for pre-processed (high-pass filtering) and transformed signals (13 time domain and frequency domain features). The LOSO cross-validation method confirmed the successful recognition of seven activities with 96.82 %. The use of two acceleration sensors, worn on the participant's ankles and wrist, was a good decision. The only drawback of the analysis could be the participant's limited age range between 20 and 25 years.

presented a classification system for six different hand gestures, where the subjects carry a sensor in the hand. They compared two sensors (shimmer sensor versus tri-axial accelerometer sensor) with 38 healthy subjects, pre-processed data, eight features and two different classifiers (SVM and KNN). Improvements were achieved upon considering both shimmer and accelerometer data. Among the classifiers considered, KNN had an average accuracy around 6 % better than SVM. Unfortunately, the LOSO validation was not considered, and the exact pre-processing strategy was not explained.

showed that online human activity recognition (1 s time windows) with one acceleration sensor mounted at the right thigh is possible for five activities with high accuracies. Even with the LOSO technique, results as high as 92 % could be achieved. Furthermore, locomotion speed was estimated for the classified activities walk and run. Overall, two SVMs fulfilled these tasks, and six young subjects participated in the study.

tried to develop a HAR system with high accuracy and low power consumption. Therefore, studies with 1 Hz sampling rate and long time windows were conducted and showed very good results for four activities and three young men. To classify two static and two dynamic activities of daily life, three SVMs had been used in a hierarchic manner to classify low-pass-, mean- and median-filtered acceleration signals. In future, more participants and real-life environments should be analysed. Drawbacks of the method are the small number of activities and the huge time windows, as the width of the time windows had to be increased when sampling rates decreased.

In general, HAR systems with wearable sensors achieve good performance. So, it is worth it to further investigate such systems to remain competitive and to produce further improvements to pave the way for future technologies.

For data collection, three different hardware systems have been used, as there have been multiple enhancements during the classification analysis. For system 1, a customized data acquisition system with acceleration sensor ADXL335 and micro-controller PIC32MX695F512H has been used. The data are transferred via LAN cable to the PC. So, the orientation and rotation of the sensor is restricted during carrying in the trouser pocket. For system 2, we used a Decawave EVK1000 evaluation kit with micro-controller STM32F105xx as base station. Additionally, a self-built tag with a LIS3DH acceleration sensor and a STM32L051x8 micro-controller was utilized. The data are transmitted via radio wave from the tag to the base station and transferred via LAN cable to the PC. System 3 is nearly the same as system 2, but instead of the LIS3DH acceleration sensor an IMU LSM9DS1 was used to record acceleration and gyroscope data. Table  shows photos of the different systems. We recorded accelerometer data with 1 kHz using system 1, accelerometer data with 50 Hz using system 2, and accelerometer and gyroscope data with 59.5 Hz using system 3. During the analysis with system 1, we concluded that even 20 Hz would be enough for good results; therefore, we reduced the sampling rate with progression of hardware development.

Within the experiments with systems 1 and 2, one healthy (self-assessed) female person was used for data generation. For analysis with system 3, 12 healthy people (5 female, 7 male) served as subjects. The subjects decided by themselves how the sensor was to be initially put in the pockets of the trousers and how fast or clear each activity was to be performed.

In our analysis, we tested various movements such as toe tip, heel tip, swivelling hips, stamping with feet, moving the knee up and down, or moving the knee left and right for sufficient recognition. Finally, we decided that the therapy table should move upwards when the masseur moves the knee up and down such as using an air pump. The toes are kept on the ground during the pumping process. Otherwise, the therapy table should move downwards, when the masseur moves the knee left and right with the toes kept on the ground. If the masseur is doing anything else, such as standing, giving a massage, going or running, the therapy table should remain in its current position.

For data collection, three classes of movements have been recorded with five different activities:

Class 1 was go, run or massage.

Furthermore, we recorded data streams (called “online data” in the next sections) with the following sequence of activities: 1.

go – class 1 for 25 s,

In the data pre-processing step, the data are segmented into windows of 3 s, and these windows are computed every 200 ms out of 10 online data streams of one subject (test data). For training, 55 samples of each kind of activity are used from a separate recording with the same subject. Therefore, class 1 consists of more training data as classes 2 and 3 as there are more different movements in it. In detail, we chose an approach with classification into five classes followed by a renaming of classes 1–5 to new classes 1–3. For the execution of the two different pumping exercises, the same leg, as where the sensor was mounted or pushed into the pocket, was used. After segmentation, multiple features are generated and normalized to values between 0 and 1.

For the acceleration data of each axis as well as acceleration magnitude, 43 features are computed, which are short-time energy (STE), short-time average zero-crossing rate (ZCR), average magnitude difference (AMD), root-mean-square energy (RMS), 15 specific band energies (BE1–BE15), spectral centroid (SCENT), median of peak differences (PEAKDIFF), number of peaks (PEAKS), spectral roll-off (SROLL), spectral slope 1 (SSLOP1), spectral slope 2 (SSLOP2), spectral spread (SSPRE), spectral skewness (SSKEW), spectral kurtosis (SKURT), spectral bandwidth (SBAND), spectral flatness (SFLAT) and the first 14 mel-frequency cepstrum coefficients (MFCC-1 to MFCC-14). In summary, we have 172 features of which 3 × 4 are of the time domain, 26 × 4 are of the frequency domain and 14 × 4 are out of the cepstrum. Further information about our used features and additional features can be found in most of the articles in our reference list .

For evaluation of different machine learning applications, we collected many features in a database that could be reasonable for such purposes. Afterwards, with feature selection or elimination processes, we automatically find a suitable subset of features. For our daily work with different classification tasks, this prefabricated feature-based approach will reduce our time for future analyses. In another publication , we also used various features to analyse vibration data of bearings. In , also a huge list of features is generated.

For classification, we used the one-nearest-neighbour (1NN) method, as in our literature study KNN methods reached classification accuracies that were as good as other methods, such as CNN. In and , the KNN method performed even better compared to other classic approaches for classification. For our application, it is important to quickly get a classification response for motor control. Therefore, we decided to use an easy and comprehensible method, and, according to the literature research, it is also comparably as good as other approaches.

Using the 1NN method, for each test data sample the most similar training data sample is searched, and the class of this training data sample is assigned to the test data sample. Additionally, a smoothing process is used to overcome the problem of individual false assignments. For this idea, a time window of length 4 s is moved over the classification results of the data stream. The most commonly estimated activity class within this period is then the new activity class for the timestamp at the middle of the time interval. This results in a time lag of 2 s for online execution until the final class is certain. In Fig. b, an example is given for the classification results of an online stream with and without smoothing. The red line shows the estimated classes without smoothing, the green line shows the estimated classes with smoothing and the blue line is the correct result.

We computed for each feature the ability to classify between the three classes. Remarkably, the features computed from acceleration magnitude, acc, and from acceleration of the z axis, accZ (accZ shows horizontally forward), performed best for the hip sensor position. The best feature (SBAND computed from acc) reached 81.42 %. For the trouser pocket sensor position, features computed from acc and accX performed best. The best feature was MFCC-4 of acc with 95.22 %. The worst features for both sensor positions reached 33.33 %.

Furthermore, we compared the two different sensor positions (hip and trouser pocket) with different sets of features:

Feature set 1 was 172 features

For our studies in this section, we used a new hardware system – see also hardware system 2 in Sect. . One subject made records with a tri-axis accelerometer with a sampling rate of 50 Hz and acceleration data limited within a range of ±2 g gravity. As in Sect. , three classes are used. Class 1 consists of 1614 data samples with activities “go”, “run” and “massage”, class 2 consists of 538 data samples for the activity “pumping the therapy table up” and class 3 consists of 538 data samples for the activity “pumping the therapy table down”. In this section, the sensor is placed into the pocket of the dominant leg (the leg that executes the key movements for classes 2 and 3). So, 50 % of the data are recorded with the sensor in the right pocket of the trousers and making the key movements for pumping the therapy table up or down with the right leg. The other 50 % of data are recorded with the sensor in the left trouser pocket and with a dominant left leg. Again, we use the 1NN method as classifier and features were computed from acceleration magnitude, acc. Due to the conducted feature engineering process, the list of features was adjusted to STE, short-time average zero-crossing rate of signal minus 1 (ZCR1), short-time average zero-crossing rate of signal minus 1.3 (ZCR2), AMD, RMS, eight specific band energies of spectrum (BE1–BE8), SCENT, median of peaks in spectrum (SPEAKS), SROLL, SSLOP1, SSLOP2, SSPRE, SSKEW, SKURT, SBAND, SFLAT and spectral flux (SFLUX). Features 1–4 are from the time domain, and features 5–25 are from the frequency domain. In this section, we did our analysis without cepstrum coefficients, as we wanted to find an easy lightweight method for classification with a low computational complexity. We set the window length to 2.88 s (144 data points) and shifted the windows every 0.32 s (16 data points). For the results with an additional smoothing process, we used only the last two classification estimates to form a new one (but with more weight for class 1). So, if there are two estimates with classes 1 and 2, we still remain in class 1. If there are two estimates with classes 2 and 3, we choose class 2 as the preferred class.

With these 25 normalized features, a forward feature selection as well as a backward feature elimination process was started for the above-mentioned offline training data of this section. These processes quickly find a local optima in contrast to brute-force methods where all parameter combinations are tested. The forward feature selection process starts with a feature, for which the maximum classification accuracy is obtained (and is easy to compute if there are more possibilities), and then successively adds a further feature, such that the classification accuracy is as high as possible .

In Fig. a and b, the progress of classification accuracy can be seen for adding features and eliminating features, respectively.

Figure a shows that with only three features a good classification can be achieved but with seven features the best result is achieved (99.87 %). These specific features are BE1, SFLAT, SFLUX, RMS, BE4, BE5 and BE6. Figure b depicts that three features are enough for a good accuracy (99.58 %) but with eight features an accuracy of 99.97 % is possible. These features are BE3, SBAND, BE2, SFLAT, BE4, SSLOP1, SKURT and BE1.

In a next step, we made a parameter optimization for these two specific parameter selections regarding window length, window shift and time lag of the smoothing process. Therefore, we used 10 online streams as test data and the offline data as training data. Interestingly, we already used an optimal combination of window length (2.88 s), shift (0.32 s) and time lag (0.32 s) for an additional smoothing process. With these parameter settings, we calculated the classification accuracies for five different feature sets, which are listed in Table . There, the eight features found via the backward elimination process achieved the highest performance.

Another interesting question is whether training data recorded with one dominant leg are enough for control of the system with both legs. The results are posted in Table  for two different feature sets. If we change the dominant leg for training and testing for the same person, the performance does not drop significantly. So, if desired, training with one leg is enough for operating the system with both legs.

During the next study, we made recordings with changes in the initial orientation of the tag in the trouser pockets as well as changes in how the base station is positioned on the table in front of the subject. We made four recordings with the four different possibilities of how the tag is placed into the pocket. Then we reversed the base station and made, again, four recordings with the four different tag orientations. All in all, the classification accuracies barely varied (between -1.13 to +0.51).

Finally, we studied how different parameter values, K, of the KNN classifier affect the classification accuracy. Up to now, we used K = 1 for hardware system 2, but now we also consider higher parameter values (K) when using eight features derived from backward elimination and how the classification accuracy will change. Therefore, we show for the best scenario of Table  the corresponding figure with alternating K. As we can see in Fig. , K = 1 or K = 3 would be a good choice. So, it practically does not matter if you choose K = 1 or K = 3.

In this section, we switched to another hardware system, denoted as hardware system 3 in Sect. , and we made recordings with 12 subjects (5 female, 7 male). The height of the participants ranged from 1.65 to 1.86 m, with the location of the sensor placed in the pockets and varied between groin and a deep position at the thigh. Also, the initial orientation of the sensor in the pockets and varied at random. Furthermore, the subjects decided which leg to use for controlling the therapy table (only restriction: leg which records the signals must be the same as the leg for control). The clarity and pace of movements also varied between the different subjects.

We recorded data with a tri-axis accelerometer and a tri-axis gyroscope with a sample rate of 59.5 Hz. As in Sects.  and ​​​​​​​, offline and online data were recorded. The recorded offline data streams (used for training) have a length of 30 s for each activity group and subject. The five activities are go, run, massage and two key movements that imitate the operation of a foot pump (knee pumps in vertical or horizontal direction). The first three activities are collected in their own class. The online data streams with various activities (used for testing) consist of three runs of a length of 2 min and 15 s for each subject as listed in Sect. . Additionally, we made records of five further possible movements for controlling the therapy table. So, the offline data streams have been extended with the following activities: stamping, toe-tipping, describing a circle with the knee, swinging the hips left and right, and quickly moving the knees forward and backward in an alternating manner.

In Table , the details for the best-reached classification accuracies of Sects. – are summarized.

The reasons for very good classification accuracies with one subject can be summarized as follows: 1.

The person is skilled, so the algorithm behind the system is known and, therefore, clear and consistent motions are beneficial.

Especially for analyses with the LOSO validation, it is known that classification accuracies are lower as with cross-validation. For instance, in and , the LOSO validation reached an accuracy of 78.35 % and 87.6 %, respectively. Our result of 86.18 % is similarly good, but please keep in mind that we computed this result out of our online data streams. For comparison purposes of our classification accuracies to others, we would like to emphasize that the accuracies in literature typically correspond to our offline data approach, where signal segments are extracted very carefully from different classes.

In this study, a HAR system carried by masseurs for controlling a therapy table via different movements of a single leg has been developed. In our experiments, we studied two different sensor positions: fixed at the right hip like a belt and loosely inserted in one pocket of the trousers. The second position turned out best, as movements are more distinct.

With one female subject, classification accuracies of about 98 % for online data streams (following a predefined protocol of movements) and up to 100 % for offline data samples (precise extraction of signal samples of distinct classes) were achieved. Thereby, three operating classes have been used: pump the therapy table up (class 2), pump the therapy table down (class 3) and do nothing (class 1) for all other activities performed by the masseur.

Furthermore, we conducted studies with 12 subjects and many different approaches and modifications. In conclusion, the classification accuracies varied in the range of 84 % to 98 % with, for instance, different validation techniques. In contrast to other literature studies, our results are comparably good.

With several subjects, the use of a minimal training database and/or a foregoing teaching process that may result in similar data for the training database may be worth consideration.

For future work, we additionally focus on key activities with a high frequency (double-clicks or tapping motions) such that the reaction time of the system can be further reduced, as the motor control cannot be faster as 1–2 cycle durations of these key motions. Moreover, we will study a special two-stage system to allow for remote control only if a special key is activated (for instance by an additional sensor or voice control).