Wearable device with a human activity recognition system and a human activity recognition method for determining the human activity of a user wearing the wearable device

Abstract

Human activity recognition (HAR) method for determining the human activity of a user wearing a wearable device (100 ) wherein such a wearable device () comprises at least one sensor (121) which is configured to detect a movement along at least one axis of a tridimensional Cartesian reference system and to generate a corresponding detection signal, a processing and control unit (122) associated to the at least one sensor (121), said HAR method comprising the steps : -receiving the detection signals x t ,...,x t–T– 1 generated by the at least one sensor (121), where T is the size of the detection time interval and t indicates a time instant; -classifying the activity of the user by executing a HAR machine learning algorithm comprising multi-class classification task T(0) wherein the execution of said multi-class classification task T(0) is carried out by executing a plurality of n sub-tasks {T (1),..., T (n)} according to a hierarchical scheme, the execution of each one of the n sub-tasks {T (1),..., T (n)} being carried out by sequentially executing a single feature extractor (FE) module and a following respective fully connected (FC) module.

Description

[0001] WEARABLE DEVICE WITH A HUMAN ACTIVITY RECOGNITION SYSTEM AND A HUMAN ACTIVITY RECOGNITION METHOD FOR DETERMINING THE HUMAN ACTIVITY OF A USER WEARING THE WEARABLE DEVICE ------ The present invention refers to a wearable device, in particular to a wearable device comprising a human activity recognition (HAR) system. The present invention refers also to a HAR method for determining the human activity of a user wearing the wearable device. Human activity recognition is a research area focused on developing systems that can automatically identify user activities, such as lying, standing, walking, or running. This task commonly utilizes Machine Learning (ML) techniques. As it is known, HAR techniques can be categorized into two main types as it is disclosed in [1], [5]. A first type comprises the so-called vision-based HAR techniques and a second type comprises the so-called sensor-based HAR techniques. The vision-based HAR techniques involve the analysis of images or videos captured by optical sensors, whereas the sensor-based HAR techniques utilize data from wearable and environmental sensors, such as accelerometers and gyroscopes as it is disclosed in

[0011] . With particular reference to the sensor-based HAR techniques, they can be divided in two sub-types; a first sub-type comprises Machine Learning (ML) based algorithms (e.g., Support Vector Machine, K-Nearest Neighbour, and Decision Trees) and a second sub-type comprises Neural Network (NN) based algorithms (e.g., Convolutional Neural networks (CNNs) and Recurrent Neural Networks (RNNs)), as it is disclosed in [1]. In

[0010] a dataset consisting of data produced from the accelerometer sensor and gyroscope sensor of a smartphone is used to show the performances of different ML algorithms. The analysis shows that CNNs are the ones providing the largest recognition abilities. Remarkably, sensor-based HAR is typically conducted over a fixed-duration window of signals and, in this perspective, the selection of the window size and sampling frequency becomes crucial when designing a HAR pipeline. In [4] it has been extensively explored how different windowing techniques impact the recognition system performance. The findings reveal that shorter windows, specifically those lasting 2 seconds or less, yield the highest accuracy in detection performance. While in [2] a study on 15 public datasets shows that a frequency of 15–20 Hz may be sufficient for HAR on a wearable (specificity / sensitivity higher than 95%). It is known that some HAR techniques use a Hierarchical approach, which, unlike the traditional methods that utilize a single model for user activity recognition, takes into account the similarities between activities and decomposes the multi-class classification problem into multiple stages of subclassification as it is disclosed in

[0022] . For example the first stage might predict a general activity category(e.g., walking or running), and subsequent stages refine the prediction by identifying specific sub-activities(e.g., walking upstairs or downstairs). In other words it is known to consider a tree-based activity recognition model that first infers the abstract activity and then identifies the specific activity in a top-down scheme. This approach improves discrimination between similar activities without explicitly assuming temporal relations between actions and activities as it is disclosed in [3], [6],

[0017] ,

[0021] ,

[0022] . In any case, the HAR methodology may be implemented by using the so-called Tiny Machine Learning (TinyML) algorithms, that are ML algorithms that can be actuated within wearable devices, constrained by limited memory, low computational power, and low power consumption as it is disclosed in

[0020] . TinyML’s popularity comes from its capability to process data locally, improving privacy, and security, reducing latency, and enabling offline operation without a constant internet connection. In TinyML, a significant focus lies on evaluating methods aiming at reducing the size and complexity of the ML models. These methods encompass precision scaling, which involves techniques like quantization (see

[0012] ) and compression(see

[0013] ). Additionally, strategies like task dropping are explored to alleviate computational burdens (see

[0019] ). Another area of exploration in TinyML involves redesigning the network architecture, such as implementing approximate 2D convolutions (see [7] and

[0023] ). Regarding on-device learning, several studies in the literature have presented techniques, targeting either specific layers or the entire architecture of Fully Connected Neural Networks (FCNN) (see

[0016] ,

[0018] ). Moreover, there have been contributions focusing on training all the layers of Convolutional Neural Networks (CNNs) (see [8]). All these approaches focused on performing on-device learning by optimizing the memory and the computational use but not on enhancing the on- device learning process with limited data availability which is a common scenario in sensor-based HAR. The object of the present invention is to overcome the above-mentioned drawbacks and in particular to devise a wearable device with a human activity recognition system and a human activity recognition (HAR) method for determining the human activity of a user wearing the wearable device, which are more computationally efficient and more accurate than the prior art. Another object of the present invention is to devise a wearable device with a human activity recognition system capable of adapting its HAR algorithm in real-time to individual users. Further another object of the present invention is to devise a wearable device with a human activity recognition system capable of adding new tasks to its HAR algorithm at runtime. These and other objects according to the present invention are achieved by making a human activity recognition (HAR) method as set forth in claim 1 and by making a wearable device with a human activity recognition system as set forth in claim 4. Further characteristics of the HAR method and of the wearable device with a human activity recognition system are the objects of the dependent claims. The characteristics and advantages of the HAR method and of the wearable device with a human activity recognition system according to the present invention will be more evident from the following exemplary though non-limiting description, referring to the attached schematic drawings in which: - Figure 1 is a hierarchical scheme of a plurality of sub-tasks of a multi-class classification task used in a HAR method according to the present invention; - Figure 2 is a scheme of the training process of the machine learning algorithm used in the HAR method according to the present invention; - Figure 3 is the hierarchical scheme of figure 1 with a new sub-task; - Figure 4a is a table showing a comparison in terms of memory, computation and latency using the STM32-NUCLEO- F401RE evaluation board; M^^^, C^^^, ^^^^ are respectively the memory and computation load of the i- th sub-task T(i) for the feature extractor (FE) module and the fully connected (FC) module; - Figure 4b is a table showing the performance comparison between hierarchical approaches related to the hierarchical scheme of figure 1; - Figure 4c is a table showing the performance comparison between a traditional single-model solution, a general known hierarchical-model solution and a hierarchical approach according to the present invention; - Figure 5 is a confusion matrix (i.e., a table comparing the predicted and actual class labels) showing the performances of the single-model approach; - Figure 6a is a graph showing the performances across various data partition percentages of learning the new task walking upstairs against walking downstairs when the present invention is not employed; - Figure 6b is a graph showing the performances across various data partition percentages of learning the new task walking upstairs against walking downstairs when the present invention is employed; - Figure 7a is a graph showing the performances across various data partition percentages of learning the new task drinking against sitting when the present invention is not employed; - Figure 7b is a graph showing the performances across various data partition percentages of learning the new task drinking against sitting when the present invention is employed; - Figure 8 is a block scheme representing a wearable device according to the present invention. With reference to the figures, a wearable device is shown, globally referred to as 100. For example, the wearable device 100 may be an eyewear, a generic head-mountable device, a smartwatch and so on. The wearable device 100 advantageously comprises a human activity recognition (HAR) system 120. The HAR system 120 comprises at least one sensor 121 which is configured to detect a movement along at least one axis of a tridimensional Cartesian reference system and to generate respective detection signal. The wearable device 100 may comprise a frame 110 or a generic structural member adapted to be worn by a user. In this case, the HAR system 120 is associated to the frame 110 and the at least one sensor 121 is coupled to the frame 110. Preferably, the at least one sensor 121 comprises one or more sensors belonging to the group comprising a monoaxial accelerometer capable of detecting accelerations along one axis of the tridimensional Cartesian reference system, a triaxial accelerometer capable of detecting accelerations along three axis of the tridimensional Cartesian reference system, a gyroscope. The HAR system 120 comprises also a processing and control unit 122 associated to the at least one sensor x. In particular the processing and control unit 122 is configured for receiving the detection signals and determining the human activity of the user wearing the wearable device 100 on the basis of the detection signals by actuating a HAR method for determining the human activity of the user wearing the wearable device according to the present invention. For example the processing and control unit 122 is an electronic processor like a microcontroller. The processing and control unit 122, in particular, can comprise an internal memory 125 wherein a computer program is loaded; such a computer program comprises instructions that, when the computer program is executed, induce the processing and control unit 122 to actuate a HAR method for determining the human activity of the user wearing the wearable device 100 according to the present invention which will be described in the following. The HAR system 120 comprises also a battery 123 associated with the processing and control unit 122, and at least one sensor 121 so as to electrically supply them. The HAR system 120 may also comprise a wireless communication module 124 associated to the processing and control unit 122 configured to communicate in a wireless manner with an electronic terminal, such as a smartphone. Preferably the processing and control unit 122 and the battery 123 are mounted in two different parts of the frame 110. However they can be also mounted inside the same part of the frame. In any case the interconnections between the at least one sensor 121, the processing and control unit 122 and the battery may be achieved by means of wires or flexible printed circuit boards PCBs. The HAR method provides the steps:- receiving the detection signals generated bythe at least one sensor 121, where T is the size of the detection time interval and t indicates a time instant; - classifying the activity of the user by executing a machine learning algorithm comprising multi-class classification task T(0)wherein the said multi-class classification task T(0)is carried out by executing aplurality of ^ sub-tasks {^^^^, … , ^^^^} according to ahierarchical scheme, each one of the ^ sub-tasks being carried out by sequentially executing asingle feature extractor (FE) module and a following respective fully connected (FC) module. The feature extractor (FE) module and the fully connected (FC) modules are software modules. In detail, each task with i ranging from 1 to n is a classifier that maps an unseen window of size ^ ofsequential data ^^ , … , ^^^^^^ to its label as follows: being:• ^^ ∈ ℝ^^^ the input of size ^^^ at time ^^^^• ^^ a label that belongs to the label set = where #^ is the total number of classesrel to the task The sequential execution of a subset of the sub-tasksset is equivalent to the execution of ^$ ^ asintroduced in

[0022] . Indeed, in the hierarchical processthe predicted label from the last executed sub-^^%^in the hierarchy belongs to the label set ^$^of ^^$^,formally: , with ^^% being the last executed sub-In Fig. 1, an example of HAR scheme with ^ = 6 ispresented. Both the subset of the sub-tasks to be executed and their execution order are determined through a hierarchical scheme. To express the dependencies between tasks, it isintroduced the matrix ' of dimension ^ × ^, representingthe dependencies between tasks as follows: ^^ ^ ^é,^ ^^ … ,^' =ê ⋮ ⋱ ⋮ùú êê,^^^^ … ,^^^ú^ úë û,^^^ ∈ { ^ ^ ∪ ∅} denotes whether the task on the task . More specifically:• = ∅ indicates the absence of dependencies of from the task indicates that a dependency exists ^Moreover, the label , of mustpredicted to trigger the activation of the task .To illustrate the concept of dependencies, consider taskshown in Figure 1. Having ^6^ = {walking,stairs} thedependency is: ,^5^6 = stairs indicating that in order for ^^5^to be executed, must predicts “stairs”. According to the present invention each one of the ^sub-tasks {^^^^, … is carried out by sequentiallyexecuting a single feature extractor (FE) module and a following respective fully connected (FC) module. For sequential execution it is meant that the output of the single feature extractor is used as the input for the respective fully connected module. This means that the single feature extractor module is executed just one time when the first sub-task is executed; the results of the execution of the single feature extractor module are stored in the internal memory 125 and used for each sub- task subsequent to the first one1 by the respective fully connected module. The execution of the single feature extractor is not repeated each time a sub-task is executed; it is the output of the first execution of such a single feature extractor that is repeatedly used at each sub-task by the respective fully connected module. In particular, all the n sub-tasks rely on the execution of the same feature extractor, for each sub-task changing just the respective fully connected module. In other words, the single feature extractor module is used for the execution of all the sub-tasks, whereas the multiple fully connected modules are used one for each specific sub-task. This approach leads to a general FE without specific task dependencies, while the FC modules are specific to their respective sub-tasks. In this way the HAR method according to the present invention reduces both the computational time and the memory required for its execution since the execution of the single feature extractor module is not repeated at the execution of each sub-task but it is run just one time at the execution of the first sub-task, the other sub-tasks involving just the execution of their respective fully connected modules using as input the results of the execution of the single feature extractor module.More formally, we refer to FE and to FCrelated to the task as ℎ^^^^⋅^. Therefore, each task being the predicted label for task at time ^. As it may be derived from Fig.2 and from the equation above reported it is clear that the single feature extractor module is run one time and the results of this execution is used by the fully connected modules of each sub-tasks. The function ;^⋅^ is implemented through a convolutional FE, a component typically consisting of ? consecutiveconvolutional blocks. Each block @ ∈ A1, ?C contains aconvolutional layer with dimensions D^E^ × D^E^and filters, which is then followed by a Max Pooling layer. Differently, the FC component ℎ^^^^⋅^ of task is designed with F dense layers, each having dense unitswith I ∈ A1, FC. The final layer is a softmax dense layerthat yields #^ outputs The machine learning algorithm is trained by a trainingprocess where ^ labels are given to each window sample^^, … , ^^^^^^. In particular, for each task a label ispicked. This is expressed more precisely as: The training process of ;^⋅^, and the ℎ^^^s is carried out simultaneously by considering the problem as a multi- output classification problem, as shown in Figure 2.Therefore, after defining a proper loss functionL^^^M^^^^, ^>^^^N for each sub-task ^^^^, the total loss functionL is defined as: with ∈ ℝR a parameter that determines whether theoptimization process should focus on minimizing the lossL^^^associated with task for the given sample.The term denotes the probability of activating the corresponding task Therefore, is implemented as follows: ^^where ∈ ℝR is the confidence level with which ^ ,2is predicted. The final layer of all FC modules exploits a softmax operation, hence allowing to consider the confidence equivalent to the softmax score for the ^corresponding label , ^^2 . It is underlined that Y^∅^ = 0.The inference is carried out directly on the device hierarchically. Specifically, ;^⋅^is executed on the input window ^^ , … , ^^^^^^, and then a subset of is sequentially executed on the output of;^⋅^ following the hierarchical schema defined in '.Specifically, the first FC module ℎ^^^^⋅^to be executed is the one without any dependency, hence the one satisfying the following condition: = ∅ ∀^ ∈ A`, aC.n, depending on the output c^_^The bd of the executed taske^_^, two conditions are possible:i. the predicted label is the algorithm final prediction and it does not trigger the execution of any other tasks: ii. the predicted label is the dependency for task thus its execution is triggered: This analysis is carried out up to when the condition (i) is not satisfied. Preferably, the HAR method comprises the steps: - receiving an indication that a new sub-task e^agh^has to be learned; - executing the HAR machine learning algorithm in order to choose the node of the hierarchical scheme to add the new sub-task e^agh^according to predetermined criteria; - executing a second learning process of the HAR machine learning algorithm to build the new sub-task e^agh^by considering the single feature extractor module and creating a new respective fully connected module; - adding the new sub-task e^agh^to the chosen node. The indication that a new sub-task e^agh^has to be learned may be generated by a command of the user, for example when the user pushes a button on the wearable device. In this way the HAR ML algorithm enters into a configuration mode in which the second learning process and the adding of the new sub-task e^agh^are executed. After the adding step the HAR ML algorithm saves the new hierarchical scheme and exits from the configuration mode. The HAR machine learning algorithm is configured to perform the on-device learning of an additional sub-task T(new)by learning a new respective fully connected module. Indeed, the feature extractor i^⋅^ is intentionally designed to be general, and without any particular task dependencies. Consequently, learning a new task T(new)allows to keep fixed i^⋅^ and fine-tune the parameters of to minimize the corresponding loss k^agh^. The memory lhrequired to store the weights of an additional task is computed as: Where q(u) being the number of dense units at layer n ∈A`, oC with b the number of dense layers.For the on-device training, the memory needed for the activations is computed as the maximum sum of memory required for the activations of two consecutive layers, as outlined in

[0015] . Another important consideration involves determining where the newly learned task e^agh^must be integrated within the hierarchical scheme established by p. This is done by using the existing scheme of the trained machine learning algorithm to predict labels for the on- device collected dataset. Subsequently, the chosen node to add e^agh^must meet the following criteria: a) it must be the most frequently predicted label among the existing ones, andb) the difference between the relative frequencies q ando of the chosen node and of the second most predictedlabel, respectively, must exceed a user-definedthreshold r ∈ As, `C. Formally: q − o > r. Therefore q >rRvw with v = q + o ≤ `.If the second condition is not met, the new task T (new) is added to both the selected node and to the second most predicted label. Figure 3 reports an example of this node selection process. Specifically, we aim to add the task of distinguishing between walking upstairs and downstairs. Given that the majority of the data corresponds to walking (with 959 samples over 1209), the walking nodeis considered a potential candidate for the new taske^agh^. Subsequently, if we set r = s. z the secondheuristic is also satisfied. Indeed, the difference in {z{frequency between the top two predicted labels is −w|}`ws{ ≈ s. z{, which exceeds r = s. z. Note that if r wereset to s. ^, the second heuristic would not be met, and T(new) would have been added to both running and walking. The HAR machine learning algorithm according to the present invention has been compared with other known algorithms by using the UCA-EHAR (University of Côte d’Azur-Embedded Human Activity Recognition) dataset reported in

[0014] which includes gyroscopic and accelerometer data collected from smart glasses worn by 20 subjects engaged in various activities, and by using the UCI-HAPT (Human Activity and Postural Transitions) dataset reported in

[0024] . In particular, the UCI-HAPT dataset is an extension of the UCI-HAR dataset

[0025] and it includes gyroscopic and accelerometer data collected from 30 people engaged in 6 basic activities, similar to the original UCI-HAR dataset with a waist-mounted smartphone; additionally, the UCI-HAPT dataset introduces 6 postural transition activities, expanding on the original activity set. As a comparison the following architectures have been employed: - a traditional single-model solution as in

[0014] having a single network where the output corresponds to the performed task. - a hierarchical solution as in

[0022] having multiple networks, one for each sub-task. To ensure a meaningful comparison, the experimental analysis employed the feature extractor module i^⋅^ and the fully connected module j^⋅^ proposed in

[0014] . Specifically, the architecture utilized a ResNetv1-6 with a single dense layer having the same number ofneurons as the number of classes to be classified (i.e.,The activities of

[0014] utilized in this analysis are: standing, sitting, walking, lying down, walking downstairs, walking upstairs, and running. Furthermore, for all the architectures, the employed hierarchical schema is the one reported in Figure 1, which has been generated as suggested in

[0022] .With a specific focus on memory usage, computation requirements, and mean latency, the findings are summarized in Table I of figure 4a. The assessment is carried out by using an STM32- NUCLEO-F401RE evaluation board, which is commonly utilized in TinyML applications. This board serves as a representative platform for evaluating the feasibility and effectiveness of the proposed architecture within the TinyML context. The results indicate that the mean latency remains comparable to that of traditional single-modelsolutions, while demonstrating a w × reduction comparedto the hierarchical ones. Additionally, similar observations apply to both memory and computationalload, with the proposed architecture exhibiting only a} KiB increase in memory compared to the single-modelsolutions, while showing a slight improvement in terms of computational load compared to the traditional single-model. In Figure 5 the confusion matrix of the non-hierarchical solution is reported. Notably, distinguishing between sitting and standing poses a challenge, as these activities are generally indistinguishable by using only the accelerometer and the gyroscope. Moreover, the single-model architecture struggles to discriminate between closely related activities such as walking, walking upstairs, and walking downstairs. This is one of the reasons that motivates the employment of a hierarchical solution. Indeed, as reported in Table II of figure 4b and Table III of figure 4c, the overall accuracy achieved for both applications (UCA-EHAR and UCI-HAPT) through the hierarchical approach surpasses that of the non-hierarchical method. For instance, from figure 4b Table II, in the case of task involving the discrimination between walking and walking onstairs, the accuracy reaches s. ^^. In contrast, thegeneral known hierarchical-model solution yieldsaccuracies below s. ^^ for the same task. The results,reported in Table II of figure 4b, indicate that the ML algorithm of the present invention performs slightly less effectively than a general hierarchical architecture (e.g., for of the ML algorithm of the present invention the F1-score is ^% less than the hierarchical solution). However, the ML algorithm of the present invention excels in tasks characterized by limited data, such as e^z^, where only 1^k samples are available, yet achieves an F1-score that is `s% higher than the hierarchical solution. Moreover, from figure 4c, the ML algorithm of the present invention achieves overall higher f1-score on both UCA-EHAR and UCI-HAPT with respect to the single model and the general hierarchical algorithms of an average 2%, showing that the model is able to work robustly on different datasets. These findings are significant, as the ML algorithm of the present invention eases the learning process in scenarios where data availability is limited. In addition it has been conducted an analysis to demonstrate the advantages of utilizing the HAR method for on-device learning of new sub-tasks during the operational life of the wearable device. To emulate the on-device incremental collection of the data, the evaluation compares the performances across various data partition percentages, starting from an initial `s% ofthe data and by increasing it incrementally by 10% up to`ss%.Moreover, two distinct conditions were considered. In the first condition, the network has no prior knowledge and both are randomly initialized. In the second condition, the present invention is employed and the network utilizes the previously learned feature extractor ;^⋅^ and learns only the second part of the network as in the HAR method. The evaluation shows that the HAR method according to the present invention ensures stable performances across various data availability conditions. In contrast, when the HAR method according to the present invention is not employed, the performances are highly dependent on data availability, therefore fewer data availability leads to a significant degradation in performance. Figure 6a and Figure 7b present two experiments of on- device learning of new tasks when the HAR method according to the present invention is employed and when it is not, showing the superiority of using the HAR method according to the present invention in conditions of limited data availability. Specifically, Figure 6a reports an example in which the two classes (i.e., walking upstairs and walking downstairs) are balanced. In contrast, Figure 7b reports an example where the two classes (i.e., drinking and sitting) are stronglyunbalanced. Indeed the class sitting has 4 × samples thanthe class drinking. From the description made, the characteristics of the wearable device and of the HAR method, object of the present invention, are clear, as are the relative advantages. Differently from traditional hierarchical approaches in which multiple training processes are needed (one for each sub-task), the ML algorithm used in the present invention has a unified training process, and its design supports the on-device learning of new activities. Indeed, the ML algorithm used in the present invention demonstrates superior performance in on-device learning, in particular under conditions of data scarcity, outperforming existing hierarchical solutions, as demonstrated by experimental results on a public- available dataset

[0014] and on a resource-constrained device, i.e., the embedded board STM32-NUCLEO-F401RE. Finally, it is clear that the wearable device and the HAR method thus conceived is susceptible of numerous modifications and variations, all of which are within the scope of the invention; moreover, all the details may be replaced by technically equivalent elements. In practice, the materials used, as well as their dimensions, can be of any type according to the technical requirements. References [1] Human Activity Recognition: A Survey. Procedia Computer Science, 155:698–703, January 2019. [2] Jose Antonio Santoyo-Ramon, Eduardo Casilari, and Jose Manuel Cano-Garcıa. A study of the influence of the sensor sampling frequency [3] Oresti Banos, Miguel Damas, Hector Pomares, Fernando Rojas, Blanca Delgado-Marquez, and Olga Valenzuela. Human activity recognition based on a sensor weighting hierarchical classifier. Soft Computing, 17(2):333–343, February 2013. [4] Oresti Banos, Juan Manuel Galvez, Miguel Damas, Alberto Guill˜ en,´ Luis Javier Herrera, Hector Pomares, and Ignacio Rojas.´ Evaluating the effects of signal segmentation on activity recognition. In Ignacio Rojas and Francisco M. Ortuno Guzman, editors,˜ International WorkConference on Bioinformatics and Biomedical Engineering, IWBBIO 2014, Granada, Spain, April 7-9, 2014, pages 759–765. Copicentro Editorial, 2014. [5] Antonio Bevilacqua, Kyle MacDonald, Aamina Rangarej, Venessa Widjaya, Brian Caulfield, and Tahar Kechadi. Human Activity Recognition with Convolutional Neural Netowrks. volume 11053, pages 541–552. 2019. arXiv:1906.01935 [cs, stat]. [6] Heeryon Cho and Sang Yoon. Divide and Conquer-Based 1D CNN Human Activity Recognition Using Test Data Sharpening. Sensors, 18(4):1055, April 2018. [7] Francois Chollet. Xception: Deep Learning with Depthwise Separable Convolutions, April 2017. arXiv:1610.02357 [cs]. [8] Michele Craighero, Davide Quarantiello, Beatrice Rossi, Diego Carrera, Pasqualina Fragneto, and Giacomo Boracchi. On-Device Personalization for Human Activity Recognition on STM32. IEEE Embedded Systems Letters, pages 1–1, 2023. [9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition, December 2015. arXiv:1512.03385 [cs].

[0010] Khimraj, Praveen Kumar Shukla, Ankit Vijayvargiya, and Rajesh Kumar. Human Activity Recognition using Accelerometer and Gyroscope Data from Smartphones. In 2020 International Conference on Emerging Trends in Communication, Control and Computing (ICONC3), pages 1– 6, February 2020.

[0011] L. Minh Dang, Kyungbok Min, Hanxiang Wang, Md. Jalil Piran, Cheol Hee Lee, and Hyeonjoon Moon. Sensor-based and vision-based human activity recognition: A comprehensive survey. Pattern Recognition, 108:107561, December 2020.

[0012] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A White Paper on Neural Network Quantization, June 2021. arXiv:2106.08295 [cs].

[0013] James O’ Neill. An Overview of Neural Network Compression, August 2020. arXiv:2006.03669 [cs, stat].

[0014] Pierre-Emmanuel Novac, Alain Pegatoquet, Benoıt Miramond, and Christophe Caquineau. UCA-EHAR: A Dataset for Human Activity Recognition with Embedded AI on Smart Glasses. Applied Sciences, 12(8):3849, January 2022.

[0015] Massimo Pavan, Armando Caltabiano, and Manuel Roveri. TinyML for UWB-radar based presence detection. In TinyML for UWB-radar based presence detection, pages 1–8, July 2022. ISSN: 2161-4407.

[0016] Massimo Pavan, Eugeniu Ostrovan, Armando Caltabiano, and Manuel Roveri. TyBox: an automatic design and code-generation toolbox for TinyML incremental on-device learning. ACM Transactions on Embedded Computing Systems, page 3604566, June 2023.

[0017] Yanglin Pu, Yongqiang Jiang, and Hai-Miao Hu. HR- HAR: A hierarchical relation representation for human activity recognition based on Wi-Fi. IET Communications, 17(1):29–44, January 2023.

[0018] Haoyu Ren, Darko Anicic, and Thomas Runkler. TinyOL: TinyML with Online-Learning on Microcontrollers, April 2021. arXiv:2103.08295 [cs, eess].

[0019] Manuel Roveri. Is Tiny Deep Learning the New Deep Learning? In Rajkumar Buyya, Susanna Munoz Hernandez, Ram Mohan Rao Kovvur, and T. Hitendra Sarma, editors, Computational Intelligence and Data Analytics, volume 142, pages 23–39. Springer Nature Singapore, Singapore, 2023.

[0020] tinyml.org. Tinyml foundation, 2023.

[0021] Aiguo Wang, Guilin Chen, Xi Wu, Li Liu, Ning An, and Chih-Yung Chang. Towards Human Activity Recognition: A Hierarchical Feature Selection Framework. Sensors, 18(11):3629, November 2018.

[0022] Aiguo Wang, Shenghui Zhao, Chundi Zheng, Huihui Chen, Li Liu, and Guilin Chen. HierHAR: Sensor-Based Data-Driven Hierarchical Human Activity Recognition. IEEE Sensors Journal, 21(3):3353–3365, February 2021.

[0023] Fisher Yu and Vladlen Koltun. Multi-Scale Context Aggregation by Dilated Convolutions, April 2016. arXiv:1511.07122 [cs].

[0024] Jorge-Luis Reyes-Ortiz, Luca Oneto, Alessandro Ghio, Albert Sama, Davide Anguita, and Xavier Parra. Human Activity Recognition on Smartphones with Awareness of Basic Activities and Postural Transitions. In Stefan Wermter, Cornelius Weber, Włodzislaw Duch, Timo Honkela, Petia Koprinkova-Hristova, Sven Magg, Gunther Palm, and Alessandro E. P. Villa, editors, Artificial Neural Networks and Machine Learning – ICANN 2014, volume 8681, pages 177–184. Springer International Publishing, Cham, 2014.

[0025] D. Anguita, Alessandro Ghio, L. Oneto, Xavier Parra, and Jorge Luis Reyes-Ortiz. A public domain dataset for human activity recognition using smartphones. In The European Symposium on Artificial Neural Networks, 2013.

Claims

CLAIMS 1) Human activity recognition (HAR) method for determining the human activity of a user wearing a wearable device (100) wherein such a wearable device () comprises at least one sensor (121) which is configured to detect a movement along at least one axis of a tridimensional Cartesian reference system and to generate a corresponding detection signal, a processing and control unit (122) associated to the at least one sensor (121), said HAR method comprising the steps:- receiving the detection signals ^^ , … , ^^^^^^ generated bythe at least one sensor (121), where T is the size of the detection time interval and t indicates a time instant; - classifying the activity of the user by executing a HAR machine learning algorithm comprising multi-class classification task T(0)wherein the execution of said multi-class classification task T(0)is carried out byexecuting a plurality of ^ sub-tasks {^^^^, … , ^^^^} accordingto a hierarchical scheme, the execution of each one ofthe ^ sub-tasksbeing carried out bysequentially executing a single feature extractor (FE) module and a following respective fully connected (FC) module. 2) Human activity recognition (HAR) method according to claim 1 wherein the machine learning algorithm is trained by a first training process where ^ labels are given toeach window sample ^^ , … , ^^^^^^, the training process ofthe single feature extractor (FE) module and of the fully connected (FC) modules is carried out simultaneously. 3) Human activity recognition (HAR) method according toone or more of the preceding claims wherein the HAR method comprises the steps: - receiving an indication that a new sub-task ^^^^E^has to be learned; - executing the HAR machine learning algorithm in order to choose the node of the hierarchical scheme to add the new sub-task ^^^^E^according to predetermined criteria; - executing a second learning process of the HAR machine learning algorithm to build the new sub-task ^^^^E^by considering the single feature extractor module and creating a new respective fully connected module; - adding the new sub-task ^^^^E^to the chosen node. 4) Wearable device (100) comprising: - a human activity recognition (HAR) system (120), wherein the HAR system (120) comprises: - at least one sensor (121) which is configured to detect a movement along at least one axis of a tridimensional Cartesian reference system and to generate respective detection signal; - a processing and control unit (122) associated to the at least one sensor (121), said processing and control unit (122) being configured for receiving the detection signals and determining the human activity of the user wearing the wearable device (100) on the basis of the detection signals by actuating a HAR method for determining the human activity of the user wearing the wearable device, said HAR method comprising the steps:- receiving the detection signalsgenerated bythe at least one sensor (121), where T is the size of the detection time interval and t indicates a time instant;- classifying the activity of the user by executing a HAR machine learning algorithm comprising multi- class classification task T(0)wherein the execution of said multi-class classification task T(0)is carried out by executing a plurality of ^ sub-tasksaccording to a hierarchical scheme, theexecution of each one of the ^ sub-tasks {^^^^, … , ^^^^}being carried out by sequentially executing a single feature extractor (FE) module and a following respective fully connected (FC) module.. 5) Wearable device (100) according to claim 4 wherein the at least one sensor (121) comprises one or more sensors belonging to the group comprising a monoaxial accelerometer capable of detecting accelerations along one axis of the tridimensional Cartesian reference system, a triaxial accelerometer capable of detecting accelerations along three axis of the tridimensional Cartesian reference system, a gyroscope. 6) Computer program loadable in a memory of an electronic processor, said computer program comprising instructions that, when the computer program is executed by the electronic processor, induce the electronic processor to actuate a HAR method for determining the human activity of the user wearing the wearable device x according to one or more of the claims from 1 to 3.

Images