A Review of Embedded Software Architectures for Multi-Sensor Wearable Devices: Sensor Fusion Techniques and Future Research Directions

Abstract

The integration of embedded software in multi-sensor wearable devices has revolutionized real-time monitoring across health, fitness, industrial, and environmental applications. This paper presents a comprehensive approach to designing and implementing embedded software architectures that enable efficient, low-power, and high-accuracy data acquisition and processing from heterogeneous sensor arrays. We explore key challenges such as synchronization of sensor data streams, real-time operating system (RTOS) integration, power management strategies, and wireless communication protocols. The reviewed framework supports modular scalability, allowing for seamless incorporation of additional sensors or features without significant system overhead. Future research directions of the embedded software include Hardware-in-the-Loop and real-world validation, on-device machine learning and edge intelligence, adaptive sensor fusion, energy harvesting and power autonomy, enhanced wireless communications and security, standardization and interoperability, as well as user-centered design and personalization. By adopting this focus, we can highlight the potential of the embedded software to support proactive decision-making and user feedback through edge-level intelligence, paving the way for next-generation wearable monitoring systems.

1. Introduction

The landscape of real-time monitoring is undergoing a significant transformation, driven by the rapid evolution of embedded software and the widespread adoption of multi-sensor wearable devices. These technologies are moving beyond simple fitness tracking to become integral tools in health and industrial safety, paving the way for a future of proactive and personalized data-driven insights. The integration of multiple sensors into a single wearable device, however, presents a complex set of challenges that require sophisticated embedded software solutions for effective real-time data acquisition and processing. Wearable devices have seamlessly integrated into daily life, revolutionizing how we interact with technology and monitor our personal data. This continuous stream of data is empowering individuals to take a more robust approach to their health and wellness while also enabling new examples in remote patient monitoring and athletic performance optimization. The ability to track vital signs, sleep patterns, and physical activity provides invaluable insights for both personal well-being and clinical assessment [1,2,3,4,5].

The drive to incorporate multiple sensors into wearable devices stems from the need for more accurate, reliable, and, contextually, aware data. While a single sensor can provide a specific metric, a fusion of data from multiple sensors offers a more holistic and robust understanding of the user’s state and interaction with the environment. For example, by combining data from an accelerometer, gyroscope, and magnetometer, it is possible to achieve a much more precise and reliable measurement of movement and orientation than with any single sensor alone. This multi-sensor approach is crucial for advanced applications such as fall detection for the elderly, detailed biomechanical analysis for athletes, and monitoring the vital signs of workers in hazardous environments. The ability to process and synthesize this complex data at the device level, or the “edge,” is a key driver for innovation in embedded software, as it enables real-time feedback and reduces latency and privacy concerns associated with cloud-based processing [1,2,3,6].

Despite the immense potential of multi-sensor wearable devices, their development and widespread adoption are hindered by several significant challenges that must be addressed through intelligent embedded software design. These challenges include the following [2,4]:

• Signal dropout: The integrity of data collected from wearable sensors can be compromised by factors such as poor skin contact, motion artifacts, and environmental interference, leading to incomplete or inaccurate data sets [5,7].
• Sensor drift and calibration: The accuracy of sensors can degrade over time, a phenomenon known as “drift,” which necessitates periodic recalibration to ensure the reliability of the collected data. This process can be unwieldy for the user and requires revolutionary software solutions for automatic and seamless calibration [8].
• Power constraints and battery life: The continuous operation of multiple sensors, coupled with the computational demands of data processing and wireless communication, places a significant strain on the limited battery capacity of small, lightweight wearable devices. Energy-efficient hardware and intelligent power management software are therefore significant for extending battery life and enhancing the user experience [3,4].
• Real-time processing demands: The aggregation and fusion of data from multiple sensors in real time require significant computational power. Embedded systems in wearable devices are typically resource-constrained, making it a challenge to perform complex data processing tasks without introducing unacceptable latency. This necessitates the development of highly efficient algorithms and software architectures that can operate within these constraints [3,4].

These challenges motivate a system-level review that jointly considers sensor fusion algorithms, embedded software architectures, communication protocols, and power management strategies under real-world wearable constraints.

This paper reviews robust solutions to the challenges of multi-sensor wearable systems by detailing several key contributions of the field. These advancements aim to provide a practical and efficient foundation for the next generation of real-time monitoring applications.

To effectively address the complexities of real-time monitoring, this paper focuses on the following noteworthy issues:

• Embedded software framework: A comprehensive embedded software framework tailored for multi-sensor data processing can allow the device to reach excellent levels of performance. The framework is designed to manage the intricacies of collecting and preparing data from various sensors in a resource-constrained environment [3,4].
• Sensor performance evaluation: A thorough evaluation of sensor performance is presented, analyzing the outputs of sensors both individually and when their data streams are fused. This analysis validates the benefits of sensor fusion in improving data accuracy and reliability [3,6].
• Real-Time synchronization and fusion: This paper details the implementation of real-time synchronization and data fusion using efficient embedded techniques. This is critical for ensuring low-latency processing and immediate feedback, which are essential for real-world applications [6].

The remainder of this paper is structured as follows: Section 2 reviews embedded system constraints in wearable devices, focusing on processing, memory, and real-time requirements. Section 3 analyzes wired and wireless communication protocols for multi-sensor wearables. Section 4 surveys sensor fusion techniques, from lightweight complementary filters to Kalman-based methods, emphasizing embedded deployment suitability. Section 5 evaluates embedded software architectures, comparing bare-metal and RTOS-based designs. Section 6 discusses power management strategies and energy-harvesting approaches for wearable systems. Section 7 addresses practical deployment challenges, including scalability, reliability, user experience, privacy, and security. Section 8 provides a comparative synthesis of the literature and identifies open research challenges, while Section 9 concludes this review and outlines future research directions.

Table 1. State-of-the-art table of systems in multi-sensors.

Reference/System	Fusion Level	Sensor Types Used	Sensor Configuration	Feature Domain	Fusion Strategy	Classifier/Algorithm	Application Area
Atallah et al. [9]	Feature-level	Accelerometer (×3)	Wrist, Waist, Ankle	Time domain	Cooperative	k-NN (k = 1)	Activity Recognition
Liu et al. [10]	Feature-level	Accelerometer	Wrist, Waist	Time and frequency	Complementary	SVM, DT	Physical Activity Estimation
Bicocchi et al. [11]	Decision-level	Accelerometer	Pocket (Mobile Phone)	Time domain	Cooperative	Instance-based, k-NN	Activity Recognition
Pärkkä et al. [12]	Feature-level	Accelerometer + Physio sensors	Wrist	Time domain	Complementary	SVM	Activity Monitoring

and

Table 2. Detailed description of multi-sensor systems.

Reference/System	Accuracy (%)	Data Window Size	Evaluation Dataset
Atallah et al. [9]	91%	1 s window, 50% overlap	Custom Dataset
Liu et al. [10]	93.2%	2 s window, 50% overlap	Custom Dataset
Bicocchi et al. [11]	~75%	3 s window	Real-Life Activity Dataset
Pärkkä et al. [12]	N/A	2–5 s window	Custom Dataset

present representative multi-sensor systems reported in the literature, focusing on fusion levels, sensor types, sensor configurations, feature domains, fusion strategies, classifiers/algorithms, application areas, and performance metrics. This comparison illustrates the diversity of design choices in wearable systems; for example, Atallah et al. [9] employ cooperative feature-level fusion across multiple accelerometers for activity recognition, while Liu et al. [10] and Bicocchi et al. [11] use different combinations of sensor types, fusion strategies, and classifiers for physical activity estimation and activity recognition. Presenting these systems side-by-side highlights variations in sensor configuration, evaluation datasets, and performance, revealing gaps in scalability and real-time applicability that systematic studies aim to address.

Table 1 and Table 2 provide a comparative overview of representative multi-sensor systems reported in the literature, detailing fusion levels, sensor types, sensor configurations, feature domains, fusion strategies, classifiers, application areas, and performance metrics such as accuracy, data window size, and evaluation dataset.

This comparison highlights the diversity of design choices in multi-sensor wearable systems; for example, Atallah et al. [9] and Liu et al. [10] use feature-level fusion with accelerometers, while Bicocchi et al. [11] employ decision-level fusion with a mobile phone sensor, illustrating trade-offs in sensor placement and algorithmic complexity. Accuracy varies across systems (e.g., 93.2% for Liu et al. [10] vs. ~75% for Bicocchi et al. [11]), reflecting differences in dataset characteristics and evaluation strategies.

Analyzing these tables reveals gaps in scalability, real-time applicability, and standardization, which underscore the importance of systematic comparative studies for informing design decisions in multi-sensor wearable platforms.

Scope and Contributions of This Review

This review focuses on embedded sensor fusion systems for wearable applications, with particular emphasis on embedded software architectures, real-time constraints, communication protocols, and power management strategies. Unlike existing reviews that primarily address algorithmic aspects of sensor fusion or application-specific case studies, this work provides a system-level perspective that integrates hardware constraints, software design choices, and deployment considerations. The main contributions of this review are as follows: (i) a comparative analysis of sensor fusion techniques with respect to embedded deployment constraints, (ii) a structured evaluation of communication and software architectures used in multi-sensor wearables, and (iii) an assessment of current technological maturity, practical limitations, and open research challenges.

This review systematically evaluates embedded software architectures and sensor fusion techniques in multi-sensor wearable devices. Unlike prior reviews that focus on individual sensor modalities or a single application domain, this work synthesizes findings across multiple studies to provide a comparative analysis of algorithmic trade-offs, energy efficiency, real-time performance, and practical deployment considerations. Key contributions of this review include the following:

• Critical comparison of sensor fusion strategies highlighting accuracy, computational cost, and suitability for different wearable applications.
• Evaluation of embedded software architectures, emphasizing scalability, maintainability, and power-aware design.
• Analysis of practical challenges, including energy management, hardware limitations, and system integration issues.
• Identification of gaps in the existing literature and recommendations for future research directions, guiding both academic and industrial practitioners.

2. Embedded System Constraints in Wearable Devices

In the current section, we elaborate and review issues related to fusion strategy, embedded software architecture, real-time data acquisition, buffering and synchronization, communication protocols and power optimization techniques.

Wearable embedded systems operate under strict constraints related to processing capability, memory availability, and real-time responsiveness. Microcontrollers used in wearable platforms typically offer limited clock frequencies and memory resources, which directly influence algorithm selection and software architecture. Real-time constraints arise from the need to sample multiple sensors synchronously, execute fusion algorithms within fixed deadlines, and ensure deterministic system behavior. These constraints strongly affect the feasibility of advanced fusion techniques and motivate lightweight implementations tailored to embedded platforms.

Processing, Memory and Real-Time Constraints

An effective real-time data pipeline is fundamental to the performance of the system, ensuring that data from heterogeneous sensors is captured, temporarily stored, and time-aligned accurately before being passed to the fusion algorithms. We note here that a structured approach that leverages the features of the RTOS is employed to manage this process efficiently and prevent data loss or timing errors [13].

The acquisition of sensor data is managed by dedicated, high-priority RTOS tasks, with a separate task assigned to each sensor. To minimize latency, data acquisition is interrupt-driven. Hardware interrupt is generated by the sensor or a timer peripheral when new data is available. This interrupt triggers the high-priority acquisition task, which reads the sensor data immediately. This ensures that data capture is prioritized over all lower-priority processes, satisfying the system’s real-time processing demands [14].

To decouple the high-frequency, time-critical data acquisition tasks from the more computationally intensive and potentially slower fusion task, a buffering mechanism is implemented. Each sensor’s acquisition task writes its incoming data into a dedicated circular buffer (or ring buffer). This First-In First-Out (FIFO) structure is highly efficient in memory usage and processing overhead. The use of buffers provides essential resilience against transient peaks in processor load, preventing data loss (signal dropout) and ensuring that the fusion algorithm always has a consistent stream of data to process [13,15].

The synchronization of sensor data streams is one of the most critical challenges. Since sensors operate at different sampling rates and on independent clocks, their data must be accurately aligned in time before fusion. A system should achieve this through two key techniques [16]:

• High-resolution time stamping: Immediately upon acquisition within the high-priority task, each data sample from every sensor is tagged with a timestamp from a common, high-resolution hardware timer. This establishes a unified time base across all sensor data streams, which is essential for the fusion algorithm to correlate measurements correctly [17].
• RTOS synchronization primitives: The RTOS is used to manage the temporal alignment of data for processing. The main data fusion task waits on an RTOS synchronization object (such as an event flag or semaphore). The individual sensor acquisition tasks signal this object after placing new, time stamped data into their respective buffers. The fusion task is only activated to run when a complete and temporally consistent set of measurements is available from all required sensors [14,17].

To illustrate the differences between various real-time data pipeline architectures,

Table 3. State-of-the-art table for real-time data pipelines.

	Aspect	Traditional (e.g., EKF Fuse)	RTOS + Ring Buffer Pipeline	Middleware (μRT, etc.)
1.	Data Capture Triggering	Polling, periodic tasks	Interrupt-driven, high-priority acquisition	ISR → topic publish (ring buffer enqueued)
2.	Buffer Structure	Fixed buffers, double-buffer	Circular FIFO per sensor	Ring buffers in publish–subscribe topics
3.	Timestamping	Often software layer	High-resolution hardware timer at acquisition task	Timestamp at publish based on message info time
4.	Synchronization Mechanism	Manual polling or heuristics	RTOS semaphores/events combining all sensors	Topic subscriber triggered when all timestamps align
5.	Resilience to Processing Load	Minimal: buffer overflows possible	Buffers decouple fusion; ISR always handles capture	Topic logic drops or reschedules based on buffer age
6.	Maturity/Usage	Widely deployed in avionics/UAVs (EKF pipelines)	Common in embedded sensor systems (FreeRTOS, VxWorks)	Emerging in µRT and DDS frameworks

compares traditional, RTOS-based, and middleware approaches across key aspects such as data capture, buffering, and synchronization.

Table 3 summarizes state-of-the-art real-time data pipeline approaches in embedded sensor systems. It compares traditional polling-based pipelines (e.g., Extended Kalman Filter (EKF fuse)), RTOS-driven ring buffer pipelines, and middleware-based approaches (e.g., µRT or DDS frameworks) across key aspects such as data capture triggering, buffer structure, time stamping, synchronization, resilience to processing load, and maturity.

The table highlights the trade-offs among simplicity, determinism, and scalability: traditional pipelines are simple but susceptible to buffer overflows; RTOS-based pipelines provide predictable, high-priority acquisition and decoupled buffers; middleware-based approaches enable flexible publish–subscribe architectures but are still emerging in embedded wearable systems. This comparison allows practitioners to select the appropriate pipeline architecture based on workload, latency, and real-time performance requirements.

Overall, these constraints favor lightweight, deterministic processing pipelines and limit the practical deployment of computationally intensive fusion and middleware solutions in wearable systems.

3. Communication Protocols for Multi-Sensor Wearables

3.1. Wired Communication Interfaces (SPI, I²C, UART)

Nowadays, a framework can rely on a tiered communication strategy to manage both the on-board interfacing with sensors and the external transmission of processed data. The selection of protocols is driven by the requirements for data throughput, power efficiency, and implementation complexity within the embedded environment.

For communication between the central microcontroller and the various sensors within the wearable device, industry-standard serial protocols are employed. The choice depends on the specific sensor’s interface and data rate requirements:

• Serial Peripheral Interface (SPI): This is a synchronous serial communication protocol used for high-throughput sensors like accelerometers and gyroscopes. Its full-duplex, high-speed capabilities ensure that data can be read from sensors with minimal latency, which is essential for real-time applications [18,19].
• Inter-Integrated Circuit (I²C): This two-wire protocol is used for sensors that require lower data bandwidth, such as magnetometers or environmental sensors. Its primary advantage is the ability to connect multiple slave devices to the same bus, reducing pin count and simplifying the hardware layout [18,20].
• Universal Asynchronous Receiver-Transmitter (UART): While primarily used for debugging and console output, UART can also be utilized for interfacing with specific modules, such as some GPS receivers, that use a serial stream for data output [18,20,21].

For transmitting the fused sensor data from the wearable device to an external entity like a smartphone or a data-logging gateway, relative systems utilize Bluetooth Low Energy (BLE). BLE is the ideal choice for embedded software applications due to its extremely low power consumption, which is a critical factor for extending the battery life of wearable devices. Its architecture is optimized for sending small, intermittent bursts of data, which perfectly aligns with the needs of real-time monitoring applications. The integration of a BLE stack, managed by a dedicated low-priority RTOS task, allows the system to wirelessly send alerts, summary statistics, or raw data streams for further analysis and user feedback [21,22,23,24,25,26,27].

Figure 1 summarizes common wired communication interfaces used in wearable sensor integration, highlighting trade-offs between throughput, wiring complexity, and power consumption.

Collectively, Figure 1 supports the selection of appropriate communication interfaces based on throughput requirements, hardware simplicity and system power constraints, ensuring reliable data exchange across the wearable monitoring platform.

Table 4. Comparative overview of sensor communication interfaces.

Protocol	Use Case	Device	Advantages	Latency/Throughput	Wiring
SPI	High-speed motion sensors	Accelerometer, gyroscope	Full-duplex, low latency	≥10 Mbps	4 wires
I2C	Mid/low bandwidth sensors	Magnetometer, temperature, pressure	Multi-slave, 2-wire simplicity	100 kbps–5 Mbps	2 wires
UART	Serial modules and debugging	GPS, console	Simple P2P, minimal hardware	≤1 Mbps	2 wires
BLE	Wireless data transmission	Gateway, smartphone	Ultra-low power, burst transfers	100 kbps–1 Mbps	Wireless

summarizes the characteristics of SPI, I²C, UART, and BLE protocols, detailing their use cases, target devices, advantages, latency/throughput, and wiring requirements. This comparison highlights the complementary roles of these protocols in multi-sensor embedded systems: SPI is suitable for high-speed motion sensors, I²C for multi-device setups, UART for debugging, and BLE for low-power wireless transmission. Understanding these trade-offs is essential for designing efficient communication interfaces in real-time wearable systems.

3.2. Wireless Communication Interfaces (BLE and Related Protocols)

Improving wireless communication and security for wearables is one of the highest priorities. Thus, the research and the adoption of advanced protocols (BLE, UWB, LoRaWAN) and strong encryption/authentication mechanisms is highly important.

The evolution of wearable embedded systems increasingly relies on advanced wireless communication and robust security mechanisms. As wearables become more sophisticated, they generate larger volumes of sensor data, often transmitted in real time for health monitoring, industrial safety, or sports analytics. Traditional low-power Bluetooth protocols and simple encryption methods may no longer meet the requirements of high-throughput, low-latency, and secure data exchange. Therefore, future research is focused on integrating next-generation wireless technologies such as Bluetooth 5.3, Ultra-Wideband (UWB), LoRaWAN, and 5G, alongside end-to-end security solutions to protect sensitive data and ensure operational reliability [4,27].

Efficient, low-latency, and long-range wireless communication is crucial for continuous, real-time wearable monitoring and remote analytics. Bluetooth 5.3 offers enhanced throughput, improved energy efficiency, and extended range compared to previous versions. Future wearables could leverage adaptive channel selection, low-energy modulation schemes, and multi-device synchronization to maintain reliable connectivity in complex environments. Research could focus on dynamic topology management, interference mitigation, and coexistence with other wireless protocols in dense environments [16].

Higher data rates, improved range, and energy efficiency allow wearables to communicate more reliably with smartphones, gateways, and cloud systems. Ultra-Wideband (UWB) is emerging as a key technology for precise localization, proximity detection, and secure communications. Future research may explore integration of UWB with wearable networks, low-power ranging protocols, and privacy-preserving localization algorithms, enabling applications such as indoor navigation, asset tracking, and social distancing alerts in crowded environments [27].

UWB enables centimeter-level localization accuracy, low-latency communication, and secure proximity sensing for wearables. LoRaWAN and other long-range, low-power wide-area network (LPWAN) protocols offer the potential for remote monitoring of health or environmental parameters without relying on local gateways or smartphones. Future studies could examine hybrid connectivity strategies, where wearables intelligently switch between short-range (Bluetooth, UWB) and long-range (LoRa, NB-IoT) communication based on context, data urgency, and energy availability [16].

Low-power, long-range networks allow for continuous monitoring in remote or industrial settings with minimal energy consumption. Security is a critical future direction. Wearables often transmit highly sensitive health or biometric data, which makes them prime targets for cyber-attacks. Research must focus on end-to-end encryption, secure boot, authentication protocols, and lightweight cryptography suitable for low-power embedded devices. Advanced methods, such as homomorphic encryption and blockchain-based authentication, could allow for secure data processing in cloud or edge systems without compromising privacy [28].

Protecting wearable data against unauthorized access and tampering is essential, especially in healthcare, finance, and industrial applications. Adaptive security mechanisms are a promising research direction. Wearables could dynamically adjust encryption levels, authentication steps, or network protocols based on risk assessment, energy availability, or user activity, ensuring both robust protection and power efficiency [29].

Security protocols that adjust dynamically can balance privacy, latency, and energy consumption in real time. Interference management and coexistence are also critical. Wearables often operate in crowded frequency bands with Wi-Fi, cellular, and other IoT devices. Future research can focus on smart spectrum sensing, dynamic frequency hopping, and cognitive radio techniques to minimize communication errors and maintain reliable data transfer [16].

Cognitive frequency selection and adaptive channel management ensure consistent connectivity in dense wireless environments. Another emerging trend is edge-assisted communication, where wearable devices offload computation to edge nodes or nearby gateways, reducing latency and power requirements. Future research may explore optimized data routing, cooperative edge-wearable processing, and compression-aware protocols to maximize efficiency without compromising performance [29].

Offloading computation to edge nodes reduces latency and energy use, and improves real-time data analysis. Integration with AI-driven network management is another promising avenue. Machine learning can predict network congestion, optimize packet scheduling, and detect anomalous communication patterns, improving both performance and security. Future work could investigate reinforcement learning for dynamic communication management and anomaly detection in multi-wearable networks [30].

Machine learning enables predictive network optimization, energy-efficient routing, and early detection of security threats. Finally, future wearables may adopt multi-protocol and multi-band architectures, allowing for seamless switching between Bluetooth, UWB, LoRaWAN, Wi-Fi, and 5G, depending on latency requirements, energy constraints, and environmental conditions. This flexibility will be essential for wearables that operate in diverse scenarios, from home-based health monitoring to industrial safety systems [31].

Flexible communication architectures allow wearables to adapt dynamically, balancing energy, latency, and reliability across various networks. In conclusion, enhanced wireless communication and security are fundamental to the future of wearable embedded systems. By integrating next-generation wireless protocols, adaptive security, AI-assisted network management, and multi-protocol architectures, wearables can achieve reliable, secure, and energy-efficient data exchange, enabling new applications in healthcare, sports, industrial monitoring, and beyond. Research in low-power secure communication, hybrid networks, and intelligent edge-assisted processing will drive the next generation of connected, trustworthy, and autonomous wearables [29].

3.3. Comparative Discussion and Design Implications

From a system design perspective, the selection of communication protocols represents a trade-off between data throughput, power consumption, scalability, and implementation complexity. Wired interfaces such as SPI and I²C are well suited for short-range, high-reliability sensor interconnections, whereas wireless protocols like BLE enable flexible data transmission to external devices at the cost of increased energy consumption and latency. The maturity of these protocols makes them reliable for commercial deployment; however, challenges remain in coordinating multiple communication interfaces while maintaining low power consumption in wearable systems.

In practice, wearable systems adopt hybrid communication architectures, combining high-speed wired links for sensing with low-power wireless transmission for external connectivity.

4. Sensor Fusion Techniques for Wearable Systems

4.1. Levels of Sensor Fusion (Data, Feature, Decision)

To address the challenges of signal dropout and sensor drift in multi-sensor wearable platforms, a robust sensor fusion strategy is essential. Sensor fusion combines data from heterogeneous sensor arrays to produce more accurate, reliable, and comprehensive information than any single sensor can provide alone. The choice of fusion algorithm is critical and represents a trade-off between estimation accuracy and the computational demands placed on the embedded system. This review considers two primary fusion strategies: the complementary filter and the Extended Kalman Filter (EKF) [32].

Sensor fusion can be performed at different abstraction levels, commonly categorized as data-level, feature-level, and decision-level fusion. Data-level fusion combines raw sensor measurements and typically provides high accuracy but requires precise synchronization and increased computational resources. Feature-level fusion operates on extracted features, reducing data dimensionality while preserving essential information. Decision-level fusion integrates outputs from independent processing modules and offers robustness at the expense of reduced granularity. The choice of fusion level significantly impacts computational load, latency, and suitability for embedded wearable platforms.

4.2. Complementary and Heuristic Filters

The complementary filter is a frequency-domain filtering technique that offers a computationally efficient method for sensor fusion, making it highly suitable for real-time applications with significant power constraints. It operates by combining two or more sensor measurements of the same physical quantity. For instance, in orientation tracking, it fuses the high-frequency data from a gyroscope (which is reliable for short-term rotational changes but prone to drift) with the low-frequency data from an accelerometer and magnetometer (which are stable over the long term but noisy). The filter applies a high-pass filter to the gyroscope data and a low-pass filter to the accelerometer/magnetometer data, effectively “complementing” each other to produce a stable and accurate orientation estimate. Its simplicity and low overhead make it an excellent choice for foundational fusion tasks [3,32].

Figure 2 illustrates the basic structure of a complementary filter used in orientation estimation.

4.3. Kalman-Based Fusion Methods

For applications demanding higher accuracy and the ability to integrate a wider array of sensor inputs, the EKF provides a more sophisticated, model-based approach. EKF is an optimal state estimator for nonlinear systems. It operates in a two-step predict-update cycle [33]:

• Prediction: The filter predicts the system’s next state based on a predefined motion model.
• Update: It corrects the predicted state using the actual measurements from the sensors, weighing the correction based on the relative uncertainty of the prediction and the measurement [34].

EKF can seamlessly fuse data from multiple sources (e.g., accelerometers, gyroscopes, magnetometers, and even physiological sensors) to estimate a single, coherent state vector. While more computationally intensive than a complementary filter, its key advantages include superior accuracy, robustness to noise, and the ability to provide an estimate of its own uncertainty. This makes it indispensable for complex tasks such as precise motion tracking and navigation in GPS-denied environments while achieving high-accuracy data processing [3,34,35].

Figure 3 illustrates the modular framework for fusing multi-sensor odometry within a robotic navigation system. The diagram is structured into four primary components that collectively enable robust and continuous state estimation.

On the left, Input Features are categorized into three types: Visual Feature Quality Metrics (e.g., key point count and reprojection error), Motion Parameters (e.g., linear and angular velocities) and Wheel Velocity Data. These inputs provide the raw observables required for motion inference.

The center of the diagram presents the Estimation Core, which employs an Error-State Kalman Filter (ESKF). This core operates through a cyclic Prediction–Observation process that facilitates Continuous State Refinement, ensuring that incremental errors are corrected in real time.

On the right, Odometry Sources are listed, including Inertial Orientation (from an IMU), Wheel-Based Motion (from encoders), and Visual Frame Alignment (from a camera). These sources are fused to produce a cohesive motion estimate.

Finally, the bottom section outlines the Process Flow, which follows a sequential loop: Initialize → Predict → Observe → Correct → Reset. This flow emphasizes the recursive nature of the filter, where each iteration reduces uncertainty and enhances localization accuracy.

Together, this figure clarifies how heterogeneous sensor data is integrated within an error-state formulation to achieve precise and reliable mobile robot pose estimation, balancing accuracy with computational efficiency in dynamic environments.

A comparative analysis of the reviewed sensor fusion techniques reveals clear trade-offs between computational cost, energy consumption, and estimation accuracy. Kalman-filter-based methods offer the highest accuracy and robust performance under nonlinear dynamics but require higher processing power and memory, which can limit adoption in low-power or resource-constrained wearable devices. Complementary and heuristic filters provide lower accuracy but demonstrate superior real-time responsiveness and energy efficiency, making them more suitable for low-power wearable systems. Cascaded or hybrid methods provide intermediate solutions, balancing accuracy, responsiveness and energy use.

Table 5. Comparative analysis of sensor fusion methods.

Fusion Method	Computational Cost	Maturity Level	Key Strength	Typical Applications	Main Limitation
Linear Complementary Filter	Low	High (widely deployed)	Low computational overhead and minimal power consumption, simple implementation	Basic orientation tracking, excellent for low-power wearables	Limited accuracy in highly dynamic or nonlinear motion
Cascaded CF ([32])	Medium	Medium	Improved drift compensation without heavy computation	Robust attitude estimation in wearables/robots, good for mid-range devices	Requires empirical tuning, limited adaptivity
Extended Kalman Filter	High	High (industry-grade)	High accuracy, uncertainty modeling, sensor redundancy handling	Navigation, motion capture, GPS-denied environments, suitable for high-end wearables	High computational complexity and increased power consumption

compares sensor fusion methods in terms of computational cost, maturity, key strengths, limitations and typical applications. The analysis shows that Kalman-filter-based methods are highly mature and widely used in industry-grade systems but present adoption barriers due to computational and energy demands. Complementary filters are also highly mature for low-power applications, with fewer integration challenges, while cascaded and hybrid methods are moderately mature, often requiring empirical tuning and careful calibration.

This critical comparison highlights practical deployment considerations, including hardware constraints, power budgets, and system scalability, and identifies research gaps: the development of adaptive or hybrid fusion techniques that combine high accuracy with low computational cost, and standardized benchmarks for evaluating trade-offs in wearable sensor fusion.

4.4. Comparative Analysis and Deployment Suitability

From an embedded deployment perspective, a comparative evaluation of fusion techniques reveals that no single approach is universally optimal for wearable applications. Complementary filters are computationally efficient and robust, making them suitable for low-power, real-time systems. In contrast, Kalman-based methods provide higher estimation accuracy but require greater computational resources and careful tuning. Deployment suitability therefore depends on application-specific requirements, including accuracy, power budget, and system complexity.

When operating independently, each sensor in the wearable array demonstrated characteristic vulnerabilities that limit its effectiveness for robust, real-time applications. For instance, accelerometers are excellent for tracking dynamic movements but are susceptible to gravitational and motion-related noise, which can corrupt the data. Gyroscopes provide stable orientation information but suffer from inherent drift over time, leading to accumulating errors in continuous monitoring scenarios. Similarly, magnetometers are useful for absolute heading but are easily distorted by nearby metallic objects or magnetic fields, making them unreliable in many real-world environments [36,37,38].

In contrast, the fused data stream, which combines inputs from these heterogeneous sensors using an EKF, improves estimation accuracy and robustness. The analysis shows that the fusion algorithm effectively compensates for the weaknesses of individual sensors. For example, the gyroscope’s short-term stability is leveraged to counteract the accelerometer’s noise, while the accelerometer and magnetometer data are used to correct the gyroscope’s long-term drift [32,35,39,40].

This synergistic relationship results in a data stream that is not only more accurate than any single sensor output but also more robust to the challenges of real-world deployment. The fused data provides a stable and accurate representation of the user’s state, essential for physiological monitoring and motion tracking applications where erroneous data could lead to incorrect assessments or alerts. The enhanced performance of the fused data stream is a clear testament to the advantages of implementing intelligent sensor fusion algorithms within the embedded software of wearable devices [41,42,43,44].

Figure 4 and Figure 5 present two complementary filtering architectures for sensor fusion and orientation estimation within the wearable monitoring system.

Together, these figures contrast the EKF’s rigorous, high-accuracy state estimation with the complementary filter’s lightweight, efficient orientation tracking, clarifying the design rationale for selecting each method based on application-specific requirements in real-time health and motion monitoring.

Table 6. State-of-the-art table of different algorithm types.

	Algorithm Type	Pros	Cons	Typical Error
1.	EKF (a + g + m)	Long-term drift correction; widely used, balance of accuracy/compute	Nonlinear model; requires tuning Jacobians	~2–4° orientation
2.	Complementary filter + bias	Computationally efficient, robust	Less flexible to complex dynamics	<3–4° peak error
3.	UKF/SR-UKF	Better nonlinear handling, more accuracy	Higher compute requirement	Slightly better accuracy
4.	Particle filter (PF)	Handles non-Gaussian noise well	Highest computational cost	Comparable

compares four commonly used sensor fusion algorithms in embedded real-time applications, highlighting trade-offs between computational complexity, accuracy, and robustness to sensor noise.

EKFs provide high accuracy and long-term drift correction but require careful tuning of model parameters and higher computational resources, making them moderately suitable for real-time embedded implementations. Complementary filters are computationally efficient and robust to transient sensor disturbances, offering excellent suitability for low-power, real-time wearable systems, albeit with slightly lower accuracy. Unscented Kalman filters (UKF/SR-UKF) improve nonlinear handling and accuracy at the cost of increased computational load, while particle filters offer strong noise rejection and robustness in non-Gaussian scenarios but impose the highest computational requirements.

This analysis illustrates that lightweight, model-based techniques (EKF, complementary filter) dominate current wearable implementations due to their balance of accuracy, robustness, and real-time viability. At the same time, more advanced methods, including UKF and particle filters, present opportunities for future development as embedded hardware continues to improve.

These figures illustrate two distinct sensor fusion frameworks for state estimation in navigation and tracking systems.

Figure 6 presents an Underwater Navigation Sensor Fusion architecture, which integrates data from primary sensors, including an Inertial Measurement Unit, acoustic positioning, velocity measurement and depth sensing, to perform error-state estimation and multi-sensor fusion. The processing pipeline transforms raw sensor inputs into corrected position and attitude estimates along with associated error bounds, emphasizing robustness in challenging environments where traditional positioning systems may be unreliable.

Figure 7 outlines a Sequential State Estimation process structured around a recursive predict–update cycle. The workflow begins with initialization of state and uncertainty, proceeds through a prediction phase based on motion modeling, and culminates in an update phase where multi-sensor measurements are integrated to refine the state estimate. The depicted sensor suite, including inertial, wheel-based, visual, and distance sensors, supports a versatile fusion strategy applicable to ground and aerial robotic platforms.

Together, these figures demonstrate complementary fusion methodologies: one tailored for underwater navigation with explicit error-state correction, and another designed for general sequential estimation with phased prediction and update operations. Both highlight the importance of structured sensor integration in achieving accurate, reliable pose estimation across diverse operational domains.

Table 7. Representative sensor fusion methods used in wearable and embedded applications.

Reference	Fusion Method	Sensor Set	Key Benefits
A Review on Multisensor Data Fusion for Wearable Health Monitoring [3]	Generic multi-sensor fusion	accelerometer, gyroscope, magnetometer, physiological sensors	Robustness to sensor faults, compensates unreliable input
Extended Kalman Filter for Real-Time Indoor Localization by Fusing WiFi and Smartphone Inertial Sensors [40]	EKF fusion	gyroscope + accelometer + magnetometer + Wi-Fi RSSI	Corrects drift, high orientation and positional accuracy
Robust Extended Kalman Filtering for Systems with Measurement Outliers [45]	EKF variant	IMU data with outlier detection	Enhanced robustness to disturbances and sensor anomalies
Particle Filtering and Sensor Fusion for Robust Heart Rate Monitoring Using Wearable Sensors [46]	Particle filter + fusion	PPG + IMU motion data	Excellent noise rejection, reliable heart-rate estimation
Multimodal Fusion for Robust Respiratory Rate Estimation in Wearable Sensing [47]	Sequential fusion	respiratory + motion sensors	Smooth output stream, reduced noise for clean physiological features

summarizes representative sensor fusion approaches reported in the literature, highlighting the diversity of fusion methods, sensor combinations, and application domains in wearable and embedded systems. The selected studies illustrate how different fusion strategies—ranging from generic multi-sensor integration to Kalman and particle filtering techniques—are employed to address challenges such as sensor noise, motion artifacts, and unreliable measurements.

A clear trend emerges toward probabilistic and model-based fusion methods, such as Extended Kalman Filters and particle filters, which provide robustness and improved estimation accuracy when sensor models and noise characteristics are well understood. At the same time, simpler fusion schemes remain attractive for wearable applications due to their lower computational requirements and ease of real-time implementation.

This comparative perspective demonstrates that current wearable systems balance robustness and accuracy against computational complexity, reinforcing the continued relevance of lightweight fusion techniques while highlighting opportunities for future work on adaptive and hybrid fusion strategies.

This comparison indicates that fusion method selection in wearables is driven less by theoretical optimality and more by embedded feasibility, power budget, and real-time determinism.

5. Embedded Software Architectures for Wearables

5.1. Bare-Metal Architectures

The choice of the underlying software architecture is a critical design decision that directly impacts a wearable system’s ability to meet real-time, low-power, and scalability requirements. In bare-metal architectures, firmware executes directly on the microcontroller hardware without the support of a commercial or open-source operating system. This approach offers minimal memory footprint and very low processor overhead, making it attractive for ultra-low-power and resource-constrained wearable devices [48,49].

However, implementing complex multi-sensor systems in a bare-metal environment presents significant challenges. Concurrency is typically managed using a super-loop structure combined with interrupt service routines, which can become difficult to maintain as system complexity increases. Ensuring precise timing, deterministic behavior, and reliable synchronization across multiple heterogeneous sensor data streams requires careful manual design and extensive testing. As a result, bare-metal implementations often suffer from limited scalability and increased risk of timing inconsistencies that are unacceptable in real-time monitoring applications [48,49,50].

5.2. RTOS-Based Architectures

To address the limitations of bare-metal designs, many multi-sensor wearable systems adopt a Real-Time Operating System (RTOS) as their software foundation. An RTOS is a lightweight operating system that provides essential services for developing complex real-time embedded applications, including task scheduling, inter-task communication, and power management support [29,49].

A key advantage of RTOS-based architectures is the availability of a preemptive, priority-based scheduler, which allows system functionality to be decomposed into independent tasks. Separate tasks can be dedicated to sensor data acquisition, sensor fusion, power management, and wireless communication, ensuring that time-critical operations meet strict real-time deadlines. In addition, RTOSs provide synchronization and communication mechanisms such as semaphores, mutexes, and message queues, which are essential for safely sharing data and synchronizing events among concurrent tasks in multi-sensor fusion systems [29,48,49,51].

The task-based nature of RTOSs also promotes modularity and scalability, enabling additional sensors or system features to be incorporated with minimal redesign. Furthermore, modern RTOSs often include hooks for power management, such as tickless idle modes, allowing the system to enter low-power states when tasks are inactive, an essential capability for extending battery life in wearable devices [29].

Figure 8 presents a three-tiered system architecture commonly employed in embedded and wearable monitoring systems. The model is organized hierarchically into User Level, System Level, and Hardware Level, each representing a distinct functional layer.

The User Level encompasses Applications and Software, which define system functionality and user interaction. This layer is supported by the System Level, responsible for Resource Management, including Task Coordination, Memory Allocation, and Device Control. The lowest tier, the Hardware Level, comprises Physical Components such as Processing Units, Storage Devices, and Peripheral Interfaces, which execute the computational and I/O operations.

This layered approach promotes modularity, scalability, and maintainability, enabling efficient management of system resources and seamless integration of additional sensors or software modules, key requirements for wearable and embedded systems.

5.3. Comparative Analysis and Long-Term Deployment Implications

Table 8. Comparative evaluation of bare-metal and RTOS-based architectures.

Aspect	Bare-Metal Super-Loop	RTOS-Based Framework
Memory footprint	Minimal (<2 KB)	Small but larger (2–10 KB)
Task ordering	Manual loop or interrupts	Preemptive, priority-based scheduler
Real-time guarantees	Manual timing, non-deterministic	Deterministic scheduling
Concurrency handling	ISR-heavy, complex scaling	Native task synchronization
Modularity	Code entangled	High modularity via tasks
Scalability	Poor beyond few sensors	High, modular task expansion
Power management	Sleep in loop, custom code	Integrated (tickless idle, sleep hooks)
Maintainability	Degrades rapidly with complexity	High for long-term development
Overall suitability	Prototyping, simple devices	Recommended for deployable multi-sensor wearables

presents a comparative evaluation of bare-metal and RTOS-based architectures. Bare-metal implementations minimize memory usage and overhead but are limited in scalability, maintainability, and real-time task management, which can constrain adoption in multi-sensor wearable systems. RTOS-based designs provide deterministic task scheduling, modular expansion, and improved fault isolation, supporting more complex deployments, but introduce additional complexity and slightly higher power consumption.

In terms of maturity, bare-metal architectures are highly mature for simple prototypes, while RTOS-based frameworks are widely deployed and considered mature for multi-sensor wearable applications. Cascaded or hybrid RTOS-lightweight approaches remain moderately mature, often requiring careful configuration and tuning for resource-constrained devices.

This analysis highlights practical deployment considerations: RTOS-based architectures are generally preferred for systems requiring extensibility, maintainability, and reliable real-time performance. However, adoption barriers include the need for developer expertise, careful memory management, and system integration effort.

Research gaps include developing lightweight RTOS frameworks optimized for ultra-low-power wearable devices, establishing standardized benchmarks for comparing architectural trade-offs, and creating guidelines for seamless migration from bare-metal to RTOS platforms in energy-constrained environments.

6. Power Management and Energy Harvesting in Wearables

6.1. Low-Power Design and Scheduling Techniques

Low-power operation in wearable multi-sensor systems is primarily achieved through software-level optimizations rather than hardware modifications. Common strategies include duty cycling of sensors and peripherals, adaptive sampling based on activity or signal dynamics, and the use of low-power operating modes supported by modern microcontrollers. By selectively disabling inactive components and reducing sampling rates during periods of low relevance, these techniques significantly reduce average power consumption while preserving essential sensing performance.

Task scheduling also plays a critical role in energy efficiency. Event-driven and priority-based scheduling allow time-critical sensing and fusion tasks to execute deterministically, while non-critical processing is deferred or executed at lower frequencies. In RTOS-based systems, tickless kernels and power-aware schedulers further minimize idle power consumption by allowing the processor to remain in sleep states for extended periods.

6.2. Energy-Harvesting Sources and Architectures

Energy harvesting and power-autonomous wearables represent an important research direction aimed at mitigating battery limitations that constrain current wearable systems. By capturing energy from the user or the surrounding environment, harvesting techniques can extend operational lifetime and reduce charging frequency, although fully self-sustaining operation remains challenging in practice.

In wearable systems, harvesting mechanisms are constrained by low available power densities, intermittent energy availability, and strict form-factor requirements. As a result, practical designs focus on harvesting sources that can be seamlessly integrated into body-worn platforms while maintaining user comfort. Future research must also address material biocompatibility, long-term durability and mechanical robustness to enable everyday use.

Mechanical energy harvesting, primarily based on piezoelectric and triboelectric generators, exploits human motion such as walking, joint movement, or body vibrations. Piezoelectric harvesters commonly employ materials such as lead zirconate titanate (PZT) or flexible polymers like polyvinylidene fluoride (PVDF). While PZT offers higher energy density, PVDF-based harvesters are more suitable for wearables due to their flexibility, durability, and biocompatibility. Architecturally, these harvesters are typically implemented as cantilever beams, stacked layers, or flexible patches integrated into footwear, straps, or textiles. However, their output is highly dependent on user activity patterns, making harvested power unpredictable.

Thermoelectric energy harvesting leverages the temperature gradient between human skin and the ambient environment using thermoelectric generators (TEGs). Wearable TEGs commonly rely on bismuth telluride (Bi₂Te₃)-based materials optimized for low-temperature gradients. Although TEGs provide continuous energy generation independent of motion, the limited temperature difference in wearable scenarios restricts achievable power levels to the microwatt range. System architectures often combine TEGs with ultra-low-power DC–DC converters and energy buffers to compensate for low voltage output.

Photovoltaic energy harvesting uses ambient or solar light through thin-film or flexible photovoltaic cells. Amorphous silicon and organic photovoltaic materials are frequently employed due to their flexibility and acceptable performance under indoor lighting conditions. While photovoltaic harvesting can deliver comparatively higher power densities in outdoor environments, its effectiveness is strongly influenced by lighting conditions and user behavior, limiting its reliability as a sole power source.

From an architectural perspective, most wearable systems adopt hybrid harvesting configurations, combining multiple energy sources (e.g., motion and thermal) with a primary battery. These architectures typically incorporate an energy management unit responsible for rectification, voltage regulation, maximum power point tracking (MPPT), and controlled energy storage.

A major component of power-autonomous wearables is intelligent energy management. Even with energy harvesting, available power is limited and fluctuating. Reported approaches rely on adaptive duty cycling and energy-aware task scheduling to allocate limited power resources efficiently. For example, high-power sensing or processing tasks could be delayed until energy reserves are sufficient, while critical alerts are always prioritized [26]. Coordinating energy harvesting, low-power hardware, and adaptive software is therefore essential to ensure continuous, efficient and reliable wearable operation.

In conclusion, energy harvesting offers a promising complement to conventional battery-powered wearable systems, particularly when combined with intelligent energy management and energy-aware algorithms. However, current harvesting technologies remain insufficient to support continuous multi-sensor operation without careful system-level optimization, highlighting the need for further advances in materials, hybrid harvesting strategies, and integrated hardware–software co-design [52].

6.3. Coordination of Multiple Energy Sources

Coordinating multiple energy sources in wearable systems introduces significant system-level challenges that extend beyond hardware design. Hybrid power architectures typically combine a rechargeable battery with one or more energy harvesting modules, requiring intelligent control strategies to ensure stable and uninterrupted operation.

At the hardware level, power-path management circuits are employed to dynamically select between harvested energy and battery power based on availability and load demand. These circuits often include priority-based switching, voltage threshold detection, and energy buffering using supercapacitors or secondary micro-batteries. Such buffering mitigates the intermittent nature of harvested energy but increases system size and complexity.

From a software perspective, energy-aware task scheduling plays a critical role. Embedded software can adapt sensing, processing, and communication workloads based on current energy availability. For example, high-power operations such as wireless transmission or complex sensor fusion may be deferred when harvested energy is insufficient, while low-power monitoring tasks remain active. This coordination requires tight integration between the power management subsystem and the RTOS scheduler.

Despite promising experimental demonstrations, most multi-source energy coordination strategies remain at a research or prototype stage. Challenges include accurate energy prediction, overhead introduced by control algorithms, and ensuring real-time guarantees under fluctuating power conditions. Consequently, while hybrid energy architectures can significantly extend device lifetime, their widespread adoption in commercial wearable systems is still limited.

Overall, effective coordination of multiple energy sources requires a co-design approach, integrating harvesting hardware, power electronics, and adaptive embedded software. Advancements in ultra-low-power control circuits and lightweight energy-aware algorithms are essential to transition these architectures from laboratory prototypes to reliable, real-world wearable deployments.

6.4. Practical Trade-Offs and Maturity Considerations

From a technology maturity standpoint, many power optimization techniques are well established and commercially deployed, whereas energy-harvesting solutions remain at an experimental or supplementary stage for most wearable applications. To address the critical challenge of limited battery life in wearable devices, several software-based power optimization strategies need to be considered. These techniques are designed to minimize energy consumption without compromising the system’s real-time monitoring capabilities. The power management strategy is deeply integrated with the RTOS and the overall software architecture [23,29,48].

The primary power-saving technique is interrupt-driven processing combined with processor sleep modes. Instead of continuously polling sensors for new data, the microcontroller is placed in a low-power sleep mode whenever possible. The RTOS automatically manages this transition when no tasks are ready to run. The processor only wakes up in response to hardware interrupts, such as a “data ready” signal from a sensor or a timer event. This approach ensures that the CPU consumes significant power only when actively processing data, drastically reducing the average power consumption [24,29,48].

Furthermore, the system employs intelligent peripheral and clock management:

• Dynamic frequency scaling: The microcontroller’s clock speed is adjusted based on the current computational demand. During intensive processing, such as running the fusion algorithm, the clock speed is maximized for performance. During periods of lower activity, the clock speed is reduced to save power [25,26,27,29,52,53,54,55].
• Peripheral gating: The software selectively powers down communication peripherals like SPI, I²C, and the BLE radio when they are not in active use. For instance, the BLE module is kept in a low-power state between transmissions, and sensor communication buses are only enabled when data is being actively sampled [29,53,56,57].

Finally, the RTOS task scheduler itself is leveraged for power efficiency. By carefully prioritizing tasks and scheduling data processing in batches rather than continuous streams, the system maximizes the amount of time the processor can remain in a deep sleep state. This minimizes the energy overhead associated with frequently waking and sleeping the CPU, contributing directly to the goal of an efficient and low-power wearable system [23,29,48,56].

Table 9. Comparative overview of power management techniques for embedded wearable systems.

Technique/Approach	System Level	Key Idea	Advantages	Limitations	Technology Maturity
Tickless RTOS Idle [29]	Software (RTOS)	Suppresses periodic OS ticks during idle periods	Significant reduction in idle power consumption, easy RTOS integration	Limited benefit under high task activity	High (commercially deployed)
Duty Cycling [57]	Software/System	Periodic activation/deactivation of sensors and peripherals	Simple to implement, effective for low-duty workloads	Can increase latency and reduce responsiveness	High
Dynamic Voltage and Frequency Scaling (DVFS/DVS) [53,54]	Hardware/Software	Adjusts CPU voltage and frequency based on workload	Large energy savings under variable load	Requires hardware support and careful timing analysis	Medium–High
Peripheral Power Gating [55,56]	Hardware/Software	Shuts down unused peripherals and buses	Reduces leakage and dynamic power	Reinitialization overhead	High
Energy-Aware Task Scheduling	Software (RTOS)	Schedules tasks based on energy constraints	Improves system-wide efficiency	Increased scheduler complexity	Medium
Event-Driven Processing	System	Activates processing only on relevant events	Minimizes unnecessary computation	Not suitable for continuous sensing	Medium
Energy Harvesting Integration [52]	System wearables	Uses ambient energy (motion, thermal, solar)	Extends operational lifetime	Intermittent and unpredictable energy	Low–Medium (experimental)
Multi-Source Power Coordination	System	Combines battery and harvested energy sources	Improved resilience and autonomy	Complex control logic	Low (research stage)

synthesizes prior studies on power optimization in embedded wearable systems, focusing on widely adopted techniques such as dynamic voltage and frequency scaling (DVFS), peripheral shutdown, duty cycling, and low-power wireless communication strategies. Each entry summarizes the application domain, core techniques, and key contributions, providing a structured comparison across software, hardware, and system-level approaches.

The comparison highlights that software-centric techniques, including tickless RTOS operation, energy-aware task scheduling, and duty cycling, are highly mature and widely deployed in commercial wearable devices, offering predictable energy savings and straightforward integration. In contrast, more aggressive approaches that combine adaptive scheduling with energy harvesting remain experimental, with adoption barriers including integration complexity, intermittent energy availability, and the need for sophisticated control algorithms.

From a deployment perspective, this synthesis reveals a gap between laboratory-demonstrated power optimization methods and solutions suitable for reliable, large-scale wearable deployment. While mature techniques are routinely integrated into current platforms, emerging strategies require further validation to ensure robustness, predictability, and long-term operational reliability.

Research gaps include developing energy management strategies that combine high efficiency with low computational overhead, integrating multiple asynchronous energy sources seamlessly, and establishing standardized evaluation benchmarks for comparing power optimization techniques in multi-sensor wearable systems.

7. Practical Challenges, Gaps and Deployment Considerations

The findings from this study highlight a significant performance gap between data streams from individual sensors and those processed through the sensor fusion framework. This section will analyze these differences, focusing on the practical implications for real-time monitoring in wearable devices.

7.1. Manufacturing and Scalability Challenges

Manufacturing and scalability present significant challenges for wearable sensor fusion systems. Component variability, calibration requirements, and production tolerances can affect system performance at scale. Furthermore, cost constraints often limit the adoption of advanced sensors or processing units, requiring trade-offs between performance and affordability.

7.2. Reliability, Calibration and Long-Term Operation

Transitioning from a theoretical model to a functional wearable system requires addressing practical challenges. A framework will be better designed with these real-world deployment aspects in mind, specifically focusing on hardware constraints, scalability, and processing performance.

• Hardware constraints: A significant challenge in designing wearable devices is the trade-off between functionality and the physical limitations of the hardware. The primary constraint is managing power consumption to ensure reasonable battery life. Typical embedded software architectures reported in the literature are designed with an emphasis on low-power operation, utilizing efficient data buffering and power-aware processing techniques to meet the strict energy budgets of embedded systems. This ensures that the device can perform continuous real-time monitoring without frequent recharging, a critical factor for user adoption [58].
• Scalability to more sensors: Such frameworks were explicitly designed for modular scalability. Software architecture allows for the straightforward integration of additional sensors without requiring a complete system overhaul. This is achieved through a modular data acquisition and processing pipeline that can accommodate new data streams with minimal overhead. This flexibility is a key advantage, as it allows the platform to be adapted for more complex future applications, such as integrating environmental sensors alongside physiological ones [30,59,60].
• Latency and throughput in embedded processing: For the system to be effective in real-time monitoring, both latency and data throughput must be optimized. Reported case studies in the literature, which involved physiological monitoring and motion tracking, indicate that the system meets the demanding real-time processing demands of these applications. The embedded software’s real-time data acquisition, buffering, and synchronization mechanisms are efficient enough to process data from the entire sensor array with minimal delay. This ensures that the system’s outputs are timely and accurately reflect the user’s current state, which is crucial for applications requiring immediate feedback or intervention [61].

Figure 9 illustrates the data flow and stakeholder communication system within the wearable health monitoring platform. The architecture is structured around three core nodes: the wearable device, which collects physiological and motion data via a local connection; the mobile device, which acts as a gateway for data aggregation and preliminary processing before transmitting it over a network connection; and the cloud platform, which provides secure storage, advanced analytics, and long-term trend analysis. Authorized stakeholders, including medical practitioners, clinical researchers, and designated caregivers, access the processed data through secure interfaces, enabling remote monitoring, clinical decision support, and personalized feedback. This integrated ecosystem supports continuous health tracking while ensuring data privacy and scalable accessibility.

Table 10. Comparative overview of wearable health-monitoring systems and their embedded system constraints.

System	Sensor Coverage	Power Constrains	Latency and Throughput	Scalability	Communication/Compute Architecture
Pantelopoulos & Bourbakis (2010) [59]	Multiple physiological sensors (ECG, SpO2)	High battery efficiency, low–power focus	Real-time vital sign detection but limited throughput benchmarks	Modular wearable prototypes	Early body area network, simple wireless
Remote Health Monitoring for Elderly (2023) [30]	Bio-sensors, environmental, actigraphy	Fog-enabled to reduce power and latency	Latency-sensitive via edge/fog—~140 ms sensing-to-actuation latency	Easily extensible gateway-based systems	IoT-edge-cloud layered architecture
Adaptive Extreme Edge Computing (2021) [58]	Neuromorphic sensor fusion capabilities	Ultra-low power, memristive/CMS architectures	Designed for minimal latency, low footprint	Supports adaptive, incremental sensor additions	On-device edge compute, minimal offloading

compares representative wearable health monitoring systems across key embedded-system dimensions, including sensor coverage, power constraints, latency and throughput requirements, scalability, and communication and computing architectures. The comparison illustrates how earlier body area network-based solutions have progressively evolved toward edge- and on-device processing paradigms to meet increasingly stringent real-time and energy efficiency demands.

The analysis highlights that modular scalability and low-latency processing are common design goals across contemporary systems, while effective power management remains a central challenge. RTOS-based architectures and edge computing approaches emerge as practical enablers for balancing real-time performance and energy efficiency, reinforcing their suitability for real-world, long-duration wearable deployments.

7.3. User Experience, Privacy and Security

Sensor fusion techniques play a central role in embedded real-time wearable systems, balancing accuracy, computational load, and responsiveness. Each method presents inherent trade-offs that affect system performance, robustness, and suitability for resource-constrained platforms.

Table 11. Comparative overview of sensor fusion techniques for embedded real-time applications.

	Method	Accuracy	Computation	Real-Time Viability	Robustness	Embedded Suitability
1.	EKF	High (model-dependent)	Moderate–High	Possibly constrained	Sensitive to model/data	Moderate
2.	Complementary Filter	Moderate	Low	Real-time viable	Robust but less accurate	Excellent
3.	Deep Learning/LSTM	Very high in complex dynamic environments	Very High	Poor for real-time embedded	Good adaptability	Limited embedded readiness

provides a comparative overview of commonly used fusion techniques, highlighting these trade-offs and their implications for wearable deployment.

Table 11 compares widely used sensor fusion techniques for embedded real-time applications across key criteria, including accuracy, computational complexity, real-time viability, robustness, and suitability for resource-constrained platforms. The comparison highlights the inherent trade-offs between estimation accuracy and computational demand, which are central to embedded system design.

This comparative analysis underscores why lightweight model-based fusion techniques dominate current wearable deployments.

The future of wearable embedded systems relies heavily on user-centered design (UCD) and human factor integration. While technical performance, such as sensor accuracy, communication reliability, and power efficiency, is essential, long-term adoption depends on usability, comfort, and user engagement. Incorporating systematic user feedback, ergonomic evaluation, and adaptive interfaces can ensure wearables meet functional requirements while enhancing satisfaction, compliance, and trust in daily life. Intelligent adaptation can further enhance comfort, usability, and engagement over time [2,62].

In conclusion, user-centered design and human factor integration are critical for wearable embedded systems. By prioritizing ergonomics, adaptive feedback, accessibility and AI-driven personalization, wearables can achieve higher adoption, sustained use and meaningful impact across healthcare, fitness, industrial and everyday applications. Technical performance alone is insufficient for successful deployment without parallel attention to usability, privacy, and long-term user acceptance [58].

8. Comparative Synthesis of the Literature

This section synthesizes the reviewed literature from a system-level perspective, comparing sensor fusion techniques, software architectures, communication protocols and power management strategies. Rather than evaluating individual technologies in isolation, it highlights cross-cutting trade-offs, maturity levels and deployment constraints, providing guidance for researchers and practitioners designing embedded wearable systems.

Wearable embedded systems face challenges in interoperability due to proprietary protocols, heterogeneous data formats, and diverse application interfaces. Middleware platforms, standardized APIs, and common data schemas provide abstraction layers that enable consistent data interpretation, cross-platform analytics, and device-agnostic applications. Integration with edge and cloud infrastructures allows for distributed computation and real-time analytics while maintaining compatibility across devices. Security and privacy frameworks, including standardized encryption, authentication, and access control, are essential to ensure data protection and trustworthiness in multi-vendor environments [21,56,59,61,63].

Semantic interoperability, ontology-based data models, and modular architectures support consistent interpretation of sensor data, enable plug-and-play devices, and allow for scalable integration into large IoT ecosystems. Compliance with standards ensures reliability and trustworthiness, particularly in regulated industries such as healthcare and industrial monitoring [59,61,63].

In conclusion, standardization and interoperability are pivotal for wearable embedded systems. Across the literature, trade-offs between computational complexity, energy efficiency, real-time performance, and usability are evident. Establishing common middleware, APIs, semantic models, and security frameworks supports multi-device collaboration, efficient data exchange, and scalable deployment. Persistent gaps include the need for standardized benchmarking, long-term reliability evaluation, and integrated optimization of energy, accuracy, and usability, providing guidance for researchers and practitioners designing future embedded wearable systems.

9. Conclusions and Future Research Directions

Successful wearable sensor fusion solutions require careful balancing of accuracy, power consumption and system complexity. Lightweight fusion algorithms, RTOS-based software architectures and mature communication protocols collectively support reliable and scalable deployments, while integration feasibility and long-term maintainability remain critical considerations.

9.1. Key Takeaways for System Designers

Effective embedded software development for wearable devices hinges on integrating robust RTOS architectures, carefully selected sensor fusion algorithms, and efficient power management. Future directions point toward on-device intelligence, leveraging adaptive algorithms and lightweight machine learning models to enable real-time analytics without compromising battery life. Co-optimizing computational performance, energy efficiency, and user-centric design is essential for transitioning wearables from passive data loggers to proactive, reliable tools across healthcare, industrial, and daily life applications.

9.2. Key Open Research Challenges and Future Directions

This section summarizes key future research directions in embedded software for multi-sensor wearable devices. It discusses key challenges and representative directions related to issues such as Hardware-in-the-Loop testing, on-device machine learning and edge intelligence, adaptive sensor fusion architectures, energy harvesting and power autonomy, enhanced wireless communication and security, standardization and interoperability, and user-centered design and human factor integration.

Hardware-in-the-Loop (HIL) testing and real-world validation provide a critical methodology for evaluating wearable embedded systems. HIL integrates physical sensors, microcontrollers, and communication modules with simulated user or environmental conditions, enabling assessment of latency, synchronization, and fault tolerance. This approach ensures that algorithms, software architectures, and adaptive fusion strategies are robust, reliable, and ready for deployment across healthcare, sports, and industrial applications [17,39,64].

On-device machine learning and edge intelligence are critical for enabling real-time activity recognition, anomaly detection, and context-aware adaptive decision-making. Lightweight ML models, optimized for embedded platforms using techniques like quantization or pruning, allow for local inference that reduces latency, preserves privacy, and supports energy-efficient operation. These methods facilitate personalized and adaptive wearables capable of interpreting and responding to user-specific physiological or activity patterns [24,43,51].

Adaptive sensor fusion architectures dynamically balance computational load, energy consumption, and accuracy by adjusting processing strategies based on sensor reliability, user activity, and environmental context. Event-driven processing and machine-learning-assisted fusion optimize resource usage while maintaining high-fidelity monitoring. These architectures also support distributed multi-node fusion for coordinated wearable networks, enhancing system robustness and scalability [26,32,45,58,60,65,66].

Collectively, future research should focus on HIL validation, on-device ML, adaptive sensor fusion, energy harvesting, wireless communication optimization, standardization, interoperability, and user-centered design. Addressing these areas will enable wearable embedded systems that are reliable, energy-efficient, context-aware, and capable of providing actionable insights in real-world applications.