Human Activity Recognition on Embedded Devices: An Edge AI

Approach

Graziele de C

assia Rodrigues

1 a

and Ricardo Augusto Rabelo Oliveira

2 b

Computing and Systems Department, Universidade Federal de Ouro Preto, Joao Monlevade, Brazil

Computing Department, Universidade Federal de Ouro Preto, Ouro Preto, Brazil

Keywords:

Human Activity Recognition, Embedded Devices, TensorFlow Lite, Real-Time, Edge AI.

Abstract:

Human Activity Recognition (HAR) is a technology aimed at identifying basic movements such as walking,

running, and staying still, with applications in sports monitoring, healthcare, and supervision of the elderly

and children. Traditionally, HAR data processing occurs in cloud servers, which presents drawbacks such as

high energy consumption, high costs, and reliance on a stable Internet connection. This study explores the fea-

sibility of implementing human activity recognition directly on embedded devices, focusing on three speciﬁc

movements: walking, jumping, and staying still. The proposal uses machine learning models implemented

with LiteRT (known as TensorFlow Lite), enabling efﬁcient execution on hardware with limited resources.

The developed proof of concept demonstrates the potential of embedded systems for real-time activity recog-

nition. This approach highlights the efﬁciency of edge AI, enabling local inferences without the need for cloud

processing.

1 INTRODUCTION

Human Activity Recognition (HAR) involves iden-

tifying basic movements such as walking, running,

jumping, and standing still (V. Sharma and Cano,

2024). This technology has applications in various

areas, including sports monitoring, hospital care, el-

derly and child supervision, among others (Shub-

ham Gupta and Deb, 2022). However, processing

these data, commonly performed on cloud servers,

poses challenges such as high energy consumption

and reliance on Internet connectivity (Bidyut Saha

and Roy, 2024).

Traditionally, machine learning algorithms, in-

cluding convolutional neural networks (CNNs), re-

current neural networks (RNNs), and long short-term

memory networks (LSTMs), have been used for hu-

man activity classiﬁcation (Waghchaware and Joshi,

2024). However, these methods require signiﬁcant

computational resources, making them challenging

to implement on resource-constrained devices. A

promising approach to overcoming these challenges is

edge computing, where data processing occurs locally

on the device, without the need for cloud transfer.

This approach offers advantages such as enhanced

https://orcid.org/0009-0008-3190-3102

https://orcid.org/0000-0001-5167-1523

privacy, reduced energy consumption, and lower la-

tency, making it particularly relevant for applications

requiring real-time responses.

In this context, tools like LiteRT, formerly known

as TensorFlow Lite, have enabled the execution of

deep learning models on mobile and embedded de-

vices. Studies have shown that it is possible to

achieve signiﬁcant reductions in model size and

computational complexity, making it feasible to use

on resource-limited devices (Diab and Rodriguez-

Villegas, 2022). However, adapting these models to

run efﬁciently in such constrained environments still

presents challenges, particularly in optimizing infer-

ence speed and maintaining accuracy.

This work explores the development of an embed-

ded device for human activity recognition, focusing

on three speciﬁc movements: walking, standing still,

and jumping. The goal is to demonstrate the feasibil-

ity of deploying machine learning models to operate

efﬁciently on resource-constrained devices.

This paper is organized as follows. Section 2

presents related work on HAR and edge computing

approaches. Section 3 describes the proposed sys-

tem, including the hardware, data collection process,

and model implementation. Section 4 discusses the

experimental results, and ﬁnally, Section 5 provides

conclusions and future directions.

Rodrigues, G. C. and Oliveira, R. A. R.

Human Activity Recognition on Embedded Devices: An Edge AI Approach.

DOI: 10.5220/0013475300003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 973-979

ISBN: 978-989-758-749-8; ISSN: 2184-4992

973

2 RELATED WORKS AND

THEORETICAL REFERENCES

This section presents a literature and theoretical re-

view on HAR, edge computing, and TensorFlow.

2.1 Human Activity Recognition (HAR)

Human activity recognition (HAR) has been exten-

sively studied due to its applicability in areas such

as healthcare and sports. Inertial sensor-based ap-

proaches (IMUs) have stood out for their ability to

capture detailed human motion data, enabling accu-

rate and real-time analysis. When combined with ma-

chine learning algorithms, these data allow human ac-

tivities to be classiﬁed with high accuracy.

The study by (da Silva et al., 2023) developed a

system using IMUs for the collection of motion data.

The work evaluated three models of recurrent neural

networks: long short-term memory (LSTM), gated re-

current unit (GRU) and simple RNN to classify hu-

man activities, demonstrating their effectiveness in

recognizing motion patterns in HAR systems.

Likewise, (Waghchaware and Joshi, 2024) con-

ducted a review of machine learning and deep learn-

ing techniques applied to human activity recognition.

Their study explores the use of data from inertial sen-

sors and images, highlighting the performance of dif-

ferent methods and providing a critical analysis of

their advantages and limitations.

In another example, (Ann-Kathrin Schalkamp and

Sandor, 2023) investigates the use of digital sensors

for human action recognition to detect changes in

movement patterns, aiming to identify diseases be-

fore clinical diagnosis. This approach highlights the

potential of HAR systems not only for activity clas-

siﬁcation, but also for health monitoring and disease

prevention. These studies underscore the relevance of

HAR research, demonstrating its application in phys-

ical activities and health-related areas.

2.2 Edge Computing

Edge computing is an approach that processes data

near the point of capture, reducing latency and re-

liance on remote connections. This technique is in-

creasingly utilized in wearables and IoT devices, en-

abling real-time analysis, enhanced data privacy, and

energy efﬁciency.

(M. S. Elbamby and Bennis, 2019) discusses the

potential of edge computing for applications requiring

high reliability and low latency, such as virtual reality

and autonomous vehicles. The study highlights how

local processing can enhance the scalability of wire-

less systems in critical scenarios.

Similarly, the work of (M. C. Silva and Oliveira,

2022) applied edge computing during the COVID-19

pandemic, presenting a proof of concept for an in-

telligent healthcare system. This system integrated

wearable biometric sensors to monitor the vital signs

of healthcare professionals, processing data locally

to generate immediate insights. The research lever-

aged data fusion, big data, and machine learning tech-

niques, demonstrating the beneﬁts of this approach in

health and safety contexts.

Meanwhile, (et al., 2024) highlight the crucial

role of edge computing in human activity recognition,

enabling devices to perform real-time inferences di-

rectly at the data capture point, without relying on

cloud processing. This approach not only improves

energy efﬁciency but also facilitates the development

of autonomous devices capable of operating in envi-

ronments with connectivity constraints.

2.3 TensorFlow and TensorFlow Lite

Machine learning (ML), a subﬁeld of artiﬁcial intel-

ligence, enables systems to learn patterns and make

predictions based on data. Traditionally, ML models

are trained and deployed in the cloud, with local de-

vices like microcontrollers used only for data collec-

tion and transmission. However, the growing demand

for real-time processing has led to frameworks like

TensorFlow Lite, which allow models to run directly

on resource-constrained devices.

TensorFlow Lite, now known as LiteRT, is de-

signed to run machine learning models on microcon-

trollers and other hardware with severe memory and

processing constraints (A. Haj-Ali and Weiser, 2020).

It enables local inferences, eliminating the need for

continuous cloud connectivity — a crucial factor for

battery-operated devices. Additionally, reducing data

transmission promotes energy efﬁciency. In Figure

1, the complete development ﬂow using TensorFlow

Lite is illustrated.

Recent studies, such as those by (S. S. Saha and

Srivastava, 2022) and (et al., 2023), address tech-

niques for optimizing models on low-capacity de-

vices. These works explore methods like model com-

pression, quantization, and strategies to maximize ef-

ﬁciency for speciﬁc tasks, such as signal classiﬁca-

tion.

Practical applications of TensorFlow Lite in hu-

man activity recognition have been presented by

(V. Sharma and Cano, 2024) and (Bidyut Saha and

Roy, 2024). These studies demonstrate the frame-

work’s effectiveness in microcontrollers for real-time

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

974

activity classiﬁcation, highlighting TinyML as a vi-

able solution for embedded devices in applications

like health monitoring and intelligent wearable de-

vices.

3 METHODS

In this section, we describe the steps and techniques

involved in developing the embedded device for hu-

man activity recognition. Figure 1 illustrates the over-

all workﬂow used with TensorFlow Lite. The pro-

cess is divided into three main stages: data collec-

tion, model training and conversion, and deployment

on the device.

Figure 1: Development ﬂow using TensorFlow Lite.

3.1 Data Collection

Movement data was collected using four wearable

devices equipped with the BNO080 inertial sensor,

which provides accelerometer, gyroscope, and mag-

netometer readings with 9 degrees of freedom. Each

device was integrated with a NodeMCU ESP-32 mi-

crocontroller, powered by a rechargeable lithium-ion

battery, ensuring portability and extended operation.

The devices were strategically positioned on dif-

ferent regions of the legs, as illustrated in Figure 2,

to maximize movement capture and enhance activity

recognition. Data transmission was carried out via

Bluetooth at a frequency of 20 Hz, corresponding to

one sample every 50 milliseconds, ensuring a high-

resolution temporal dataset.

The dataset was collected from these four devices,

each recording accelerometer and gyroscope data for

three speciﬁc activities: walking, jumping, and stand-

ing still. Each sample comprised six normalized fea-

tures (ax, ay, az, gx, gy, gz).

The recorded activities, as described in Table 1,

include walking, standing still, and jumping. Data

collection was conducted under various conditions to

ensure the diversity and quality of the dataset. Raw

data underwent preprocessing, including noise re-

moval and normalization, ensuring compatibility with

the machine learning model.

During data processing, each device recorded ap-

proximately 1,000 samples for each gesture. The col-

lected data, sampled at a frequency of 20 Hz (1 sam-

ple every 50 ms), were grouped such that for each sec-

ond of gesture duration, 20 samples were obtained.

This procedure was applied individually to the data

collected by each device.

Figure 2: Positions of the wearable devices.

Table 1: Description of recorded activities.

Label Description

Walking Data collected indoors during a

walking activity

Jumping Data collected indoors during

jumping activities

Standing Data collected while standing still

Figures 3 and 4 display 250 samples of accelera-

tion and gyroscope data for the activities of walking,

standing still, and jumping. In the acceleration graph,

walking shows signiﬁcant variations across all three

axes with typical cyclical oscillations. When standing

still, the readings are stable and close to zero, with

minor ﬂuctuations. During jumping, the data exhibit

peaks, especially on the y-axis, and more abrupt vari-

ations due to the intensity of the movement.

In the gyroscope graph, moderate variations in an-

gular velocity during walking reﬂect the swinging of

the legs, while at rest, angular velocity remains close

to zero. During jumping, higher intensity peaks are

observed but are less frequent than in the acceleration

data.

3.2 Model Training and Conversion

In this stage, the ﬁrst step was to implement a neu-

ral network model using the TensorFlow/Keras library

to classify the actions of walking, standing still, and

jumping based on six features per sample (ax, ay, az,

Human Activity Recognition on Embedded Devices: An Edge AI Approach

975

Figure 3: Collected acceleration data.

Figure 4: Collected gyroscope data.

gx, gy, gz). The model, shown in Figure 5, is deﬁned

as a sequential network composed of Dense Layers

with ReLU activation, dropouts, and an output layer

with Softmax activation.

Figure 5: TensorFlow Model Structure.

The collected data was labeled using a one-hot

encoding scheme, where each class was represented

as a binary vector. The segmentation was performed

by dividing the continuous sensor readings into non-

overlapping windows of 20 samples per gesture, cor-

responding to 1 second of data per instance. This seg-

mentation ensured that each instance captured a full

second of motion data, preserving the temporal char-

acteristics necessary for classiﬁcation.

To ensure unbiased training, the dataset was ran-

domly shufﬂed before splitting into 60% for training,

20% for validation, and 20% for testing.

The model was compiled using the Adam opti-

mizer, an efﬁcient choice for neural networks. The

loss function employed was categorical crossentropy.

During training, the metrics monitored included accu-

racy and mean absolute error (MAE). The model was

trained for up to 500 epochs with a batch size of 32,

using the training and validation datasets.

Next, the model was evaluated to validate its per-

formance with test data. Subsequently, it was con-

verted into the TensorFlow Lite (TFLite) format for

use in embedded devices. The model was then trans-

formed into a byte array in C format using the xxd

tool. The resulting data was stored in a header ﬁle

named model.h, which can be directly integrated into

the ﬁrmware of embedded systems. The ﬁnal model

had a size of 71316 bytes. Figure 6 illustrates the

steps followed in this process.

Figure 6: Model Training and Conversion Steps.

3.3 Deployment on the Device

During the deployment stage, the Arduino Nano 33

BLE board was used, as illustrated in Figure 7. This

board was designed to cater to the growing audience

of developers, makers, and enthusiasts in the ﬁelds of

AIoT (Artiﬁcial Intelligence and Internet of Things),

combining integrated sensors and Bluetooth Low En-

ergy (BLE) connectivity. Equipped with an Arm®

Cortex®-M4F processor (with FPU) operating at 64

MHz and 1 MB of Flash memory + 256 kB of RAM,

the Arduino Nano 33 BLE is a complete and low-

power board.

Figure 7: Arduino Nano 33 BLE.

To test the feasibility of the project, sensor data

was emulated. The model was deployed on the board

using the .h ﬁle generated in the previous stage. Using

the emulated sensor features, such as acceleration on

the X, Y, and Z axes, as well as gyroscope data on the

same axes, the board was able to infer, locally and ac-

curately, the type of movement performed, as shown

in Figure 8.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

976

Figure 8: Experiment.

This experiment serves as a proof of concept and

demonstrates the potential of the device to function as

an edge system. It can receive sensor data and make

decisions about the type of movement without need-

ing to send the data to external servers. This enables

more efﬁcient operation, lower power consumption,

and greater privacy—essential characteristics for the

development of intelligent edge systems.

3.4 Evaluation Metrics

To assess the performance of the model and the em-

bedded device, several metrics were used during both

the training phase and testing with real data. Below,

the main metrics are presented

3.4.1 During Training

• Accuracy. The ratio of correct predictions to the

total number of samples evaluated.

• Validation Loss (Val Loss). Represents the aver-

age error of the model on the validation dataset.

A low value indicates that the model is ﬁtting the

data correctly.

• MAE (Mean Absolute Error). The average of

the absolute differences between predicted and ac-

tual values.

3.4.2 In the Test Data

• Accuracy. The percentage of correct predictions.

Accuracy =

TP + TN

TP + TN + FP + FN

(1)

• Precision. The proportion of correctly predicted

positive instances.

Precision =

TP + FP

(2)

• Recall. The proportion of actual positive in-

stances correctly identiﬁed.

Recall =

TP + FN

(3)

• F1-Score. The harmonic mean of precision and

recall.

F1 = 2 ·

Precision · Recall

Precision + Recall

(4)

• Confusion Matrix. A distribution of correct and

incorrect predictions, highlighting potential mis-

classiﬁcations among classes.

3.4.3 In the Application

• Inference Latency. The time required to make a

prediction from sensor data.

• Power Consumption. Measurement of energy

consumption during inference.

Power (mWs) = current (mA) · voltage (V) · time (s)

(5)

4 RESULTS AND DISCUSSION

In this section, we present the results obtained during

the development of the model and its deployment on

the embedded device.

4.1 Model Performance During

Training and Validation

Table 2 presents the metrics obtained during the train-

ing of the model, designed to recognize three human

actions: walking, jumping and standing.

The results indicate that the model achieved good

accuracy during training (93.65%) but a decline in

validation accuracy (84.78%). This difference, along

with the increase in loss and MAE, suggests early

signs of overﬁtting. As a solution, it would be per-

tinent to investigate whether the dataset is sufﬁciently

diverse or to apply training adjustments, such as reg-

ularization techniques or data augmentation.

Human Activity Recognition on Embedded Devices: An Edge AI Approach

977

Table 2: Metrics training.

Metric Training Validation

Accuracy 0.9365 0.8478

Val Loss 0.3340 0.5964

MAE 0.1120 0.1566

4.2 Evaluation on Test Data

Table 3 presents the global and per-class metrics

based on the test data and the Figure 9 as confusion

matrix, considering the classes: 0 as Walking, 1 as

Jumping, and 2 as Standing.

Table 3: Metrics on test data.

Class Accuracy Precision Recall F1-Score

Walking 0.861 0.794 0.861 0.826

Jumping 0.631 0.857 0.631 0.727

Standing 0.971 0.918 0.971 0.944

Global 0.855 0.856 0.855 0.851

Figure 9: Confusion Matrix on Test Data.

The overall results indicate that the model

achieved an accuracy of 85.56%, demonstrating its

ability to correctly classify most samples. Further-

more, the global precision of 85.63% and global recall

of 85.56% reﬂect a balance between the reliability of

positive predictions and the correct identiﬁcation of

classes. The global F1-Score, which combines these

metrics, was 85.15%, highlighting the model’s overall

consistency.

Analyzing the per-class results:

• Walking: This class showed good performance,

with an accuracy of 86.11% and high recall, indi-

cating that most examples were correctly classi-

ﬁed. However, the moderate precision of

79.49% suggests some predictions for this class

were misclassiﬁed as others.

• Jumping: This class showed the poorest perfor-

mance, with an accuracy of 63.16% and low re-

call, indicating the model struggled to identify this

class correctly. Despite this, the high precision

(85.7%) suggests that the model made some cor-

rect predictions for this class but struggled to dif-

ferentiate it from Walking.

• Standing Still: This class exhibited excellent per-

formance across all metrics, with an accuracy of

97.14%, high recall, and high precision, demon-

strating the model’s reliability in identifying this

class.

Although the model performs well overall, spe-

ciﬁc improvements can be made to increase accuracy

and recall for the Jumping class, ensuring a more bal-

anced classiﬁcation across all categories.

4.3 Evaluation During Deployment on

the Embedded Device

Table 4 presents the latency and energy consumption

values during inference.

Table 4: Metrics during inference.

Metric Result

Latency 3.2 ms

Energy Consumption 0.1056 mWs

These results indicate that executing the Machine

Learning model developed with TensorFlow directly

on the device (local inference) is highly efﬁcient in

terms of time and energy. The latency of only 3.2

ms demonstrates extremely fast processing, suitable

for real-time applications such as motion detection.

The energy consumption of 0.1056 mJ per inference

is extremely low, reinforcing the feasibility of embed-

ded device-based solutions for edge scenarios, such as

wearables and IoT.

It is important to note that the latency of the sys-

tem is also inﬂuenced by the sensor update rate. For

instance, the sensor sampling rate, often measured in

Hz, directly impacts the frequency at which data is

acquired. A higher sampling rate could result in more

frequent sensor readings, thus potentially increasing

the data throughput and the need for faster process-

ing times. Conversely, a lower sampling rate might

reduce the load on the system but could compromise

the real-time responsiveness of applications, such as

motion detection, which require high-frequency data.

Therefore, the choice of sensor update rate is a key

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

978

factor in achieving an optimal balance between per-

formance and energy efﬁciency in embedded systems.

5 CONCLUSIONS

This study demonstrates the feasibility of developing

a human activity recognition (HAR) model using Ten-

sorFlow and deploying it directly on an embedded de-

vice. The results highlight the practical application of

machine learning at the edge, providing real-time per-

formance and energy efﬁciency.

The process, which involved data collection, pre-

processing, model training, and deployment, proved

effective in recognizing three primary activities:

walking, standing still, and jumping. The model

achieved good accuracy during the training and vali-

dation phases, despite minor challenges such as over-

ﬁtting and reduced performance in speciﬁc activity

classiﬁcations, like jumping.

Deployment on the Arduino Nano 33 BLE val-

idated the model’s ability to perform accurate local

inference with minimal latency and energy consump-

tion. This makes it suitable for use in wearable de-

vices and IoT systems, where privacy, efﬁciency, and

low-power operation are crucial.

Future improvements could include diversifying

the dataset to increase representativeness, optimizing

the model’s performance for classes with lower ac-

curacy, and adding new activities to enhance the sys-

tem’s versatility. Additionally, in this study, sensor

data on the Arduino Nano 33 BLE were emulated. As

a next step, the device used for data collection could

be integrated to send information directly to the Ar-

duino Nano 33 BLE, eliminating emulation and cre-

ating a complete system.

Nevertheless, this proof of concept demonstrates

the tremendous potential of embedded systems for

real-time human activity recognition, highlighting the

efﬁciency of edge AI solutions. The study establishes

a solid foundation for future advances in HAR appli-

cations, meeting demands for privacy, low power con-

sumption, and high efﬁciency.

ACKNOWLEDGEMENTS

The authors acknowledge the use of ChatGPT, an AI

language model by OpenAI, for assisting with the

translation of this article from Portuguese to English.

REFERENCES

A. Haj-Ali, N. P. and Weiser, U. (2020). Neural networks

on microcontrollers: Saving memory at inference via

operator reordering. In 2020 ACM/IEEE 47th Annual

International Symposium on Computer Architecture

(ISCA), pages 1105–1117, Valencia, Spain.

Ann-Kathrin Schalkamp, Kathryn J. Peall, N. A. H. and

Sandor, C. (2023). Wearable movement-tracking data

identify parkinson’s disease years before clinical di-

agnosis. Nature Medicine.

Bidyut Saha, Riya Samanta, S. G. and Roy, R. B. (2024).

Towards sustainable personalized on-device human

activity recognition with tinyml and cloud-enabled

auto deployment. arXiv preprint, 2409.00093.

da Silva et al., J. C. F. (2023). Ai-based personalized

human activity recognition in walking and trekking

sports: A case study. IEEE Latin America Transac-

tions, 21(8):874–881.

Diab, M. S. and Rodriguez-Villegas, E. (2022). Embedded

machine learning using microcontrollers in wearable

and ambulatory systems for health and care applica-

tions: A review. IEEE Access, 10:98450–98474.

et al., A. M. H. (2024). Tinyml empowered transfer learning

on the edge. IEEE Open Journal of the Communica-

tions Society, 5:1656–1672.

et al., Y. A. (2023). A comprehensive survey on tinyml.

IEEE Access, 11:96892–96922.

M. C. Silva, A. G. C. Bianchi, S. P. R. J. S. S. and Oliveira,

R. A. R. (2022). Edge computing smart healthcare

cooperative architecture for covid-19 medical facili-

ties. IEEE Latin America Transactions, 20(10):2229–

2235.

M. S. Elbamby, C. Perfecto, C.-F. L. J. P. S. S. X. C. and

Bennis, M. (2019). Wireless edge computing with la-

tency and reliability guarantees. Proceedings of the

IEEE, 107(8):1717–1737.

S. S. Saha, S. S. S. and Srivastava, M. (2022). Machine

learning for microcontroller-class hardware: A re-

view. IEEE Sensors Journal, 22(22):21362–21390.

Shubham Gupta, Sweta Jain, B. R. and Deb, A. (2022). A

tinyml approach to human activity recognition. Jour-

nal of Physics: Conference Series, 2273.

V. Sharma, D. P. and Cano, J. (2024). Efﬁcient tiny machine

learning for human activity recognition on low-power

edge devices. pages 85–90.

Waghchaware, S. and Joshi, R. (2024). Machine learning

and deep learning models for human activity recogni-

tion in security and surveillance: a review. Knowledge

and Information Systems, 66:4405–4436.

APPENDIX

In this appendix, the source code for the wearable

project can be found on GitHub: GitHub - Calc¸a

Wearable Project.

Human Activity Recognition on Embedded Devices: An Edge AI Approach

979