New Framework for Human Activity Recognition for Wearable Gait Rehabilitation Systems

Abstract

This paper presents a novel Human Activity Recognition (HAR) framework using wearable sensors, specifically targeting applications in gait rehabilitation and assistive robots. The new methodology includes the usage of an open-source dataset. This dataset includes surface electromyography (sEMG) and inertial measurement units (IMUs) signals for the lower limb of 22 healthy subjects. Several activities of daily living (ADLs) were included, such as walking, stairs up/down and ramp walking. A new framework for signal conditioning, denoising, filtering, feature extraction and activity classification is proposed. After testing several signal conditioning approaches, such as Wavelet transform (WT), Principal Component Analysis (PCA) and Empirical Mode Decomposition (EMD), an autocepstrum analysis (ACA)-based approach is chosen. Such a complex and effective approach enables the usage of supervised classifiers like K-nearest neighbor (KNN), neural networks (NN) and random forest (RF). The random forest classifier has shown the best results with an accuracy of 97.63% for EMG signals extracted from the soleus muscle. Additionally, RF has shown the best results for IMU signals with 98.52%. These results emphasize the potential of the new framework of wearable HAR systems in gait rehabilitation, paving the way for real-time implementation in lower limb assistive devices.

1. Introduction

Lower limb exoskeletons represent a significant advancement in rehabilitation engineering and assistive technology, serving as wearable robotic devices that can augment, restore, or enhance human locomotor function. These devices have become increasingly important in addressing mobility challenges for individuals with spinal cord injuries, stroke, and other neurological conditions [1]. Lower limb exoskeletons can be broadly categorized into three types: medical exoskeletons for rehabilitation and mobility assistance, military exoskeletons for load-carrying and endurance enhancement, and industrial exoskeletons for worker support and injury prevention [2].

In medical applications, these devices demonstrate promise in gait rehabilitation, showing improved outcomes in walking speed, balance, and functional independence among patients [3]. Recent technological developments have led to lighter, more energy-efficient designs incorporating advanced control strategies like EMG-based systems and adaptive algorithms, making them more practical for daily use [4]. The applications of these devices extend beyond medical rehabilitation to include performance enhancement in military operations, support for industrial workers in physically demanding tasks, and assistance for elderly individuals in maintaining mobility and independence [5,6].

Researchers use different types of vision-based and sensor-based input data for HAR for data collection. Although many research works have talked about the advantages of sensor-based data compared with vision-based data, most state-of-the-art studies still use video cameras (i.e., vision-based) for HAR due to their high accuracy [7]. Vision-based data collection approaches can be classified into two types: videos and images. For the videos in the HAR literature, collected from CCTV or smartphone devices are used, but for vision-based HAR, social media and camera images are used. On the other hand, mobile and wearable body sensors are the two types of sensor-based data sources that are found in the existing literature. Despite such importance, vision-based data are larger in size and take more processing than sensor-based data. One feasible option is increasing data density through sensors; however, while the cost of sensors has come down significantly over time compared to the alternatives (e.g., vision-based data capturing devices), it remains much more expensive. However, the computability of the body sensor system limits the execution of such a complex algorithm. Thus, sensor-based data are somehow advantageous.

HAR is the human activity detection and identification challenge in different state-of-the-art techniques [8]. Activities are mainly Activities of Daily Living (ADLs) like walking, jogging, going up or down the stairs or ramp walking. Data availability and its nature are essential parts of a reliable HAR. It has been developed as an essential research domain in ubiquitous computing and human–computer interaction, offering significant implications for healthcare monitoring, assisted living, and rehabilitation applications. The integration of wearable sensors, particularly inertial measurement units (IMUs) comprising accelerometers, gyroscopes, and magnetometers, has revolutionized the ability to accurately detect, classify, and monitor human movements in real-time environments [9].

Modern HAR systems leverage these sensor networks to capture complex motion patterns and physiological signals, enabling the detection of both basic activities (walking, sitting, standing) and more intricate movements, as can be seen in Figure 1 [10].

The advancement of machine learning and deep learning techniques has substantially improved the robustness and accuracy of activity recognition systems [11]. The authors of [12] proposed methods for identifying human activities based on a decision tree classifier. However, the classification accuracy rate is considered unsatisfactory. Cheng et al. [13] proposed three distinct classification methods, such as hidden Markov model, support vector machine, and artificial neural network, to categorize body activities. While these methods deliver acceptable performance, they are either constrained in handling significant intraclass variations or hindered by the complexity of adjusting model parameters. Furthermore, the integration of contextual information and multi-modal sensor fusion techniques has enhanced the system’s ability to distinguish between similar activities and detect transitions between different movement states, making HAR systems increasingly reliable for real-world applications.

Furthermore, Mekruksavanich et al. have introduced a novel deep learning classifier for gym activities named CNN-ResBiGRU. They collected raw EMG and IMU signals from 10 healthy subjects and achieved a classification accuracy of 97.29% [14]. Zhu et al. [15] have introduced a load-free hand rehabilitation system based on virtual reality (VR) made from ionic hydrogels. The system can identify 14 hand gestures with an accuracy of 97.9%. Another activity recognition system is developed by Lu et al. [16]. As they have produced a 5G Narrowband Internet of Things (NB-IoT) system, it is developed for human healthcare data collection, transmission, and reproduction together. The system is integrated with a bionic crack-spring fiber sensor (CSFS) inspired by Cirrus and Spider Structures. This system is characterized by its high sensitivity and long sensing range.

Another study is presented by Mengarelli et al. [17], this study investigates the feasibility of estimating the vertical component of the ground reaction force (VGRF) using only EMG signals from the thigh and shank muscles. Two deep learning models were used across three experimental setups. The findings demonstrate that EMG signals can be effectively leveraged to estimate VGRF during walking. Tigrini et al. [18] has proposed a new phasor-based feature extraction approach (PHASOR) that captures spatial myoelectric features to improve the performance of LDA and SVM in gait phase recognition. A publicly available dataset was used to evaluate PHASOR. Additionally, data-driven deep learning architectures, such as Rocket and Mini-Rocket, were included for comparison.

Moreover, myoelectric activity of muscles was used to estimate ankle kinematics as proposed by Mobarak et al. [19]. sEMG signals were recorded for a total of 288 gait cycles. Two feature sets were extracted from sEMG signals in the time domain (TD) and wavelet (WT) and compared. Then, they were used for feeding three machine learning models (artificial neural networks, random forest, and least squares support vector machine (LS-SVM)).

However, the usage of such highly complex and high processing-based classifiers is considered a costly process and needs a lot of raw data for developing a dependable classifier. At the same time, accurate classification and signal conditioning of raw and complex data such as EMG and IMU signals are a necessity. Normally, open issues and challenges that still exist in previous works related to HAR can be categorized into five categories: data collection, data pre-processing, hardware/sensors used, complex activity discovery, and non-overlap between activities.

The continuous evolution of sensor technology, coupled with sophisticated data processing algorithms, has made HAR an indispensable tool for applications ranging from fall detection in elderly care to performance analysis in sports science [20]. Recently, to achieve more flexibility in data recording, smartphones and wearables such as wristbands and smartwatches are integrated with sensors to improve the flexibility of data recording further. Such devices were used for daily health and sport users. These HAR systems can also be used for assistive and biomedical devices. For controlling an exoskeleton or artificial limb, EMGs or IMUs are typically the most convenient sensors that can be used for HAR systems. But to utilize these two types, typically a complete framework of data collection, preprocessing, feature extraction and activity classification is utilized.

For the usage of EMGs, the authors in research [21] proposed a data acquisition system for measuring EMG signals for human lower limb activity recognition. Five leg activities have been accomplished to measure EMG signals from two lower limb muscles to validate the developed hardware. Five subjects were chosen to acquire EMG signals during these activities. The raw EMG signal was first denoised using a hybrid of Wavelet Decomposition with Ensemble Empirical Mode Decomposition (WD-EEMD) approach to classify the recorded EMG dataset. Then, eight time domain (TD) features were extracted using the overlapping windowing technique.

Another example of a complete frame for activity recognition is proposed in [22] with IMU sensors. The authors have collected data from four classes of body movement datasets, namely stand-up, sit-down, run, and walk. Wearable inertial measurement unit (IMU) sensors were used for sensing and data sampling of human activity. Then, data pre-processing and feature analysis were performed by PCA and the minimum redundancy–maximum relevance (mRMR) feature selection algorithm. Finally, activity recognition was performed by traditional machine learning, deep neural networks, transfer learning and hyperparameter optimization methods. Hence, in our research, the idea of presenting a complete framework for activity recognition for assistive devices is sought after.

The first step is to provide biomechanical data from the HAR sensors. Normally, open-source datasets are used for the verification of classification and signal conditioning techniques. This can be considered the first step to producing real-time experimental hardware for exoskeleton mechanisms. Several open-source datasets were collected for HAR applications.

Moore et al. [23] presented a dataset with 15 healthy subjects. They were four females and eleven males with an average age of 24 ± 4 years, height of 1.75 ± 0.09 m, and mass of 74 ± 13 kg. The recorded activities are walking at three different speeds (0.8 m s⁻¹, 1.2 m s⁻¹ and 1.6 m s⁻¹). A total of approximately 1.5 h of normal walking and 6 h of perturbed walking are included in this dataset. The trials were performed on an R-Mill treadmill, which has dual six-degree-of-freedom force plates and independent belts for each foot. A USB 6255 card was used as a data acquisition unit. Four ADXL330 three-axis accelerometers were used as wearable sensors. An Osprey camera is used as the motion capture system.

Another open-source dataset is presented by Hu et al. [24], in which data were selected from 10 healthy subjects. There were seven males and three females. Their biometrics were 25.5 ± 2 years; 174 ± 12 cm; and 70 ± 14 kg for age, height, and weight, respectively.

The utilized sensors were sEMG, IMUs and goniometers. The sEMG (DE2.1 Delsys) was fixed on the following seven muscles in each leg: tibialis anterior (TA), soleus (SOL), vastus lateralis (VL), medial gastrocnemius (MG), rectus femoris (RF), semitendinosus (ST), and biceps femoris (BF). These signals were amplified by 1000×, band-pass filtered between 20 and 450 Hz and sampled at 1 kHz. Additionally, a 6-DOF inertial measurement unit named MPU 9250 IMUs was placed bilaterally on the subjects’ thigh (below RF) and shank (adjacent to TA) and sampled at 500 Hz.

All data were recorded by a 16-bit DAQ unit. The performed activities were a complete circuit of sitting (S), LW, ascending/descending a ramp with a 10° slope (RA/RD), standing (St), and ascending/descending a four-step staircase (SA/SD) step-over-step. A larger dataset was presented by Lencioni et al. [25], where data were collected from 50 healthy subjects. There were 25 males and 25 females. Their age range, mass and height were 6–72 years, 18.2–110 kg, and 116.6–187.5 cm, respectively. An eight-channel wireless sEMG (ZeroWirePlus) was used. Their signals were band-pass filtered at 10–400 Hz and sampled at 800 Hz, 960 Hz and 1000 Hz. They were applied on the following muscles: tibialis anterior (TA), gastrocnemius medialis (GM), soleus (SO), rectus femoris (RF), peroneus longus (PL), vastus medialis (VM), gluteus maximus (GMax) and biceps femoris (BF). Additional utilized sensors were a 9-camera motion capture system (SMART system) and two force plates (Kistler). The performed activities were walking at different speeds, toe-walking (T), heel-walking (H), step ascending (U) and step descending (D).

Another dataset was presented by Schreiber et al. [26] with 50 healthy subjects. There were 26 males and 24 females. Their age range, height and weight were 37.0 ± 13.6 years, 1.74 ± 0.09 m, and 71.0 ± 12.3 kg, respectively. A 10-camera optoelectronic system (OQUS4, Qualisys) is used for data acquisition, and the data were sampled at 100 Hz. Ground forces and moments were recorded using two force plates (OR6-5-AMTI). Their data were sampled at 1500 Hz. Eight wireless versions of sEMG (Desktop DTS—Noraxon) were used to collect muscle data from the right leg, and the utilized muscles were gluteus maximus, gluteus medius, vastus medialis, rectus femoris, gastrocnemius medialis, semitendinosus, soleus, and tibialis anterior. Bandpass filtering was applied to these data between 30 and 300 Hz. During a single session, the following exercises were carried out: walking at five different speeds on a level, straight walkway. These speeds were 0–0.4 m. s⁻¹, 0.4–0.8 m. s⁻¹, and 0.8–1.2 m. s⁻¹ in addition to other faster speeds. In total, 1143 trials were completed for all subjects and all activities.

From previous literature, it was proven the necessity of developing data classification and signal conditioning techniques and testing them on open-source datasets. In this way, new algorithms can be tested on previously validated data. Consequently, the new techniques can be applied to new experimental datasets and to human subjects. This paper presents a new methodology for the classification of ADLs using sEMG and IMUs with the intention of achieving high accuracy and low-speed classification.

In this work, our objective is to present a novel autocepstrum-based framework for studying lower limb locomotion. The remarkable characteristic of autocepstrum analysis is enhancing the significant features representing a specific activity while suppressing noise such as additive Gaussian noise according to homomorphic filtering capabilities. The proposed work captures and extracts information from different lower limb muscles to accurately recognize human movement. Indeed, many transfemoral amputees have had their lower limbs removed entirely below the knee because of illness or an accident; our proposed work is greatly inspired by this fact. An open-source dataset has been selected for testing the proposed approach to reduce the complexity of hardware preparations and facilitate algorithm testing. From our observation, wearing many sEMG or IMU sensors may make the wearer uncomfortable and require a lot of data processing and hinder the portability of assistive devices. Hence, we aim to choose between the employment of sEMG and IMU based on the obtained classification accuracy. To ensure recognition accuracy, the number of sensors must be kept to a minimum. This can be fulfilled by deciding on the muscles that have the most effective contribution in identifying a specific activity.

The methodology of this work is presented in Section 2. Different signal conditioning techniques are presented and applied to IMU and EMG signals in Section 3. The main proposed framework for activity recognition is shown in Section 4. The obtained results and their discussion and analysis are presented in Section 5. The paper is finalized with the conclusion and future work in Section 6.

2. Methodology of Work

2.1. Dataset Overview

Open-source datasets facilitate the development of new prediction techniques and provide a benchmark for making comparisons. The highly citable Georgia Tech. dataset is chosen for data collection, processing, and classification in this research. It is presented by Camargo et al. in 2021 [27]. The recorded activities are the main common activities required by elders and impaired patients in their Activities of Daily Living (ADLs), and they are used for identification of locomotion activities and gait cycle data for lower limb assistive devices [28]. Data were collected from 22 healthy subjects. There were 13 males and 9 females. Their age, height and weight are 21 ± 3.4 years, 1.70 ± 0.07 m, and 68.3 ± 10.83 kg, respectively. The utilized sensors were 3 goniometers (Biometrics), four 6-axis inertial measurement units (Yost 3-space embedded IMUs), and 32 motion capture markers with a motion capture system (Vicon). The ground reaction forces were recorded by 2 force plates (Bertec). Additionally, 11 versions of sEMG (Biometric sEMG) were used for collecting muscle data. This dataset has been a benchmark for developing and comparing conventional and deep learning techniques, as can be seen in [29,30].

The data utilized in this research is mainly from IMUs and EMGs. The IMUs were placed on the thigh, shank, trunk, and foot. On the other hand, 8 versions of sEMG were placed on the muscles of both legs. The intended muscles are gastrocnemius medialis, soleus, tibialis anterior, vastus lateralis, vastus medialis, rectus femoris, biceps femoris, gracilis, gluteus medius, semitendinosus, and right external oblique. The sensors are located on the human body as shown in Figure 2. Using a low-pass filter with a cutoff frequency of 100 Hz, the IMU data were processed after being sampled at 200 Hz. The bandpass filter was used to process the 1000 Hz sampled EMG data, with a cutoff frequency ranging from 20 Hz to 400 Hz.

2.2. EMG Signal Analysis

Dynamic electromyography (EMG) is a useful tool available to directly measure muscle activity [31]. Because the myoelectric signal well matches with the level of muscle activation, it can be a good indicator of what mechanical impact it has. EMG signals collected during gait can also be interpreted as an index of the magnitude of muscle activation. The linear envelope of the EMG signal seems to reflect the proportionate, or relative, amount of tension in muscle. These facts lead to the conclusion that technique and physiological factors play a crucial role in this relationship. Therefore, detecting alterations in the phasing, duration, or magnitude of muscle action to a pathological gait profile of an individual from such a complex EMG record is very hard to achieve. There is the multi-spike, random amplitude quality in EMG signals that makes it difficult for interpretation. This information gives a sense of the timing and intensity of muscle activity during a phase or of the whole gait cycle.

There are two key points in capturing and recording the EMG signal. The first one is the signal-to-noise ratio, which is the ratio of energy in an EMG signal to the noise signal’s energy. Normally, noise means electrical signals that are not part of the desired EMG signal. The other factor is the distortion of the signal, as the ratio of any given frequency component in an EMG signal should not be changed. All the muscle fiber action potentials from a single motor neuron together form what is called the motor unit action potential (MUAP). This is a distinctive, oscillating signal that can be detected using a non-invasive skin-surface electrode placed near the source or using an invasive electrode inserted on location into individual muscles. Equation (1) provides an overview of the structure and timing of electrical signals in muscles, especially EMG signals, as follows:

$$x\left(n\right)={\Sigma }_{m=0}^{N-1}h\left(m\right)\xi \left(n-m\right)+w\left(n\right),$$

(1)

where $x\left(n\right)$ is the modeled EMG signal, $\xi \left(n\right)$ is the point processed representing the firing impulse, $h\left(m\right)$ represents the MUAP, $w\left(n\right)$ is the zero-mean additive white Gaussian noise and $N$ is the number of motor unit firings.

EMG signals are noisy due to the tissues they travel through and motion artifact inherent noise in electronic equipment, ambient noise and motion artifacts. Many methods have been suggested for computing transitions from muscle on- and off-timing. To detect a certain activity of one patient using an EMG signal, it is necessary to analyze the data that has been obtained. In this term, burst and silence moments in electromyography may be determined as shown in Figure 3. The burst part merged with the silence part forming a full signal.

3. Signal Processing Techniques

The EMG signal must be decomposed to understand how muscle and nerve control works. Several approaches have been developed for the decomposition of EMG. To decompose the EMG signal, wavelet spectrum matching is used to calculate the use of principal component analysis of wavelet coefficients. Phinyomark et al. investigated the effectiveness of using multi-level wavelet decomposition for extracting key features from electromyography (EMG) signals [32]. Various mother wavelets and decomposition levels were tested to isolate optimal resolution components for signal reconstruction, effectively eliminating noise and irrelevant signal parts. Key features such as mean absolute value and root mean square were extracted from the reconstructed signals to enhance class separability.

The process for the multi-unit EMG signal decomposition algorithm comprises four sub-processes: signal de-noising sub-process, spike detection sub-process, spike classification sub-process, and spike separation sub-process. Daniel et al. showed that the wavelet coefficients of lower bands were more relevant in the ability to distinguish action potential (AP) characteristics than wavelet coefficients of higher bands, as shown here for various AP features [33]. Nevertheless, the authors of [34] showed experimentally that high-frequency information must be considered in the classification of MUAP. To surpass the subjective criterion for selecting the appropriate features, they suggested another approach by applying PCA on wavelet coefficients. Their approach consists of four processing stages: segmentation, wavelet transform, PCA, and a decomposition algorithm based on clustering.

The advantage of this approach is that it takes into consideration all frequency information. In [35], the authors proposed the use of nonlinear least mean square (LMS) optimization to decompose higher-order cumulants of EMG signals. Their terms have a decomposition in the third-order cumulants, multiplicative factors that show up as coefficients in nonlinear equations of motion. It is also solved in a nonlinear LMS sense, as shown next. This method adopted a multiple-input multiple-output model type due to its ability to model several MUAPs simultaneously overlapped on the EMG signal. Raw EMG signals include information grouped in random patterns. To extract this information, raw EMG signals must undergo diverse signal processing techniques. Hereinafter, a brief overview of EMG signal processing techniques is provided.

3.1. Multiresolution Analysis by Wavelet Transform for EMG Signal

Wavelet transform is one of the most effective signal processing algorithms for electromyography (EMG) signal analysis. The EMG identification system makes extensive use of it. Multiple-level wavelet decomposition is utilized to extract the EMG features [34,35]. The useful resolution components from the EMG signal were extracted using different levels of different mother wavelets. Noise and unwanted EMG parts can be eliminated through wavelet denoising. The main challenge is to choose the most suitable mother wavelet that suits the hidden target features.

3.2. PCA-Based Dimensionality Reduction

PCA has become extremely crucial for dimensionality reduction. It is used for the abstractions of fewer features depending upon the number of original signals. This approach aims at dimensionality reduction while maintaining the information of the original data. Nevertheless, the direct approach to PCA is based on eigenvalue decomposition, which is computationally demanding and shallow due to the singularity that hits when the dimension is greater than the training examples. Figure 4 and Figure 5 show examples of employing multivariate PCA for EMG and IMU signal analysis, respectively.

3.3. Empirical Mode Decomposition Approach

The EMD method employs mechanical oscillators, each of which generally only vibrates with one frequency but does so differently in another weight and tension environment [36]. To find these functions intrinsically, an algorithm thus becomes necessary; an intrinsic mode function (IMF) is a function that meets both of the following two conditions:

• In the whole dataset, the sums at which maxima or minima occur must be equal; or if this is not true, then their difference can have at most one extremum.
• For any point on that curve, its general height between two neighboring extrema and between two neighboring minima should always average to zero. Figure 6 shows a flowchart for performing averaging for EMD, which is called Ensemble EMD (EEMD).

In the literature, the use of EMD for human activity recognition includes denoising purposes and extracting features based on the chosen IMF, such as the mean absolute average, root mean square, variance, and higher-order statistics. Figure 7 represents an example of denoising an EMG before using EEMD. Figure 8 illustrates that the third IMF is the most informative component. EMD is suited for decomposing nonlinear signals such as sEMG and IMU signals. Ensemble EMD (EEMD) is used to decompose nonstationary signals. It is a noise-assisted data processing technique that evaluates the ensemble average of IMF; however, its computational complexity makes it challenging to employ in real-time applications.

3.4. Cepstral Analysis for HAR

The concept of cepstrum has been utilized in various applications such as speech recognition, gear/machine diagnostics, and echo detection/removal [37].

This analysis involves calculating the power spectrum of the logarithm of the power spectrum, with the aim of identifying echoes in seismic signals. In this manuscript, we propose the employment of the autocepstrum analysis (ACA) as a robust feature extraction approach for lower limb activity detection. Autocepstrum analysis was used for the detection of spread spectrum signals in low signal-to-noise ratio (SNR) environments [38]. The literature suggests using CA to take advantage of the discriminative activity information found in acceleration signals for HAR, which is based on homomorphic analysis. Information regarding whole-body dynamics can be separated out and converted into a compact representation known as cepstral coefficients via homomorphic analysis [39,40].

4. Proposed Approach for Lower Limb Activity Detection

Collected data from each wearable sensor must be preprocessed to eliminate artifacts before extracting. In this regard, we propose the hierarchy shown in Figure 9 for analyzing lower limb signals.

The proposed approach involves the following phases for activity detection:

• Phase I: preprocessing, which includes filtering, rectification, and denoising.
• Phase II: signal segmentation and decomposition by ACA.
• Phase III: extracting features from the autocepstrum signal.
• Phase IV: choosing a reliable classifier to decide on type of activity.

4.1. Preprocessing Phase

EMG and IMU signals acquire noise while traveling through different tissues. Some of these artifacts can be eliminated by means of high-pass filtering (rejecting certain frequency regions), such as inherent noise in electronic equipment and inherent instability due to the firing of motor units. On the other hand, some artifacts require detailed signal decomposition to separate noisy and unwanted components from the significant signal features and cannot be removed by filtering. Wavelet denoising is one of the most powerful tools to separate noisy wavelet coefficients from detailed coefficients representing features of the signal of interest. A common example of artifacts requiring denoising is motion artifacts.

4.1.1. Filtering

Applying a band-pass filter, which retains frequencies within the designated range and eliminates frequencies outside of it, is one of the most used methods for filtering EMG data. Our raw data are run through a band-pass filter at 10–400 Hz using the following functions. One of the easiest and most straightforward ways to improve the fidelity of the sEMG signal is to filter as much noise as possible while keeping as much of the required EMG signal frequency spectrum as feasible. This is in addition to employing efficient techniques for identifying and attaching the sEMG sensor to the skin. Figure 10a,b illustrates an example of filtering raw EMG that represents fast walking activity and filtering raw IMU signal, respectively. Commonly used sensors can record sEMG signals with a frequency spectrum ranging from 0 to 400 Hz, depending on the electrode spacing, the quantity of fatty tissue between muscle and skin, and the geometries of the action potentials. To preserve the desired information from the sEMG signal, band-pass filtering always strikes a compromise between reducing noise and artifact contamination.

The sensor’s bandwidth is typically higher if it is positioned above the muscle’s innervation zone or where the muscle fibers insert into the tendons. When the amplitude of the noise components exceeds that of the sEMG signal, the low-pass filter corner frequency should be placed near the high-frequency end of the sEMG signal spectrum. There should therefore be a low-pass corner frequency in the 400–450 Hz range at the top end of the sEMG frequency spectrum.

4.1.2. Denoising

The signal obtained during the detection and collecting phases might be influenced by a variety of circumstances. To improve the quality of the signal, the denoising phase is added to the preprocessing step before employing the decomposition technique. In general, a variety of noise types are generated, such as ambient noise from the human body and noise from the apparatus that records muscle signals. A high-pass filter can reduce both types of noise since they are generated by random variables. The irregular firing rate of the motor causes the signal to become unstable, which leads to the final component.

The noise frequency components are approximately 0–19 Hz. Since combined noise and artifacts cannot be eliminated by filtering, the current method denoises the sensory signals using a discrete wavelet transform [41]. For instance, orthogonal Meyer wavelets and Daubechies wavelets are frequently used to lower noise in biological signals. The chosen wavelets usually have profiles that resemble the forms of action potentials from motor units. Based on wavelet analysis, wavelets break down signals into several time-scale components [42].

From the literature, wavelets typically employed for denoising biological signals include the Daubechies (db2, db8, and db6) wavelets and orthogonal Meyer wavelet [43]. Typically, wavelets that resemble the MUAP in shape are selected. The resulting discrete wavelet coefficients are then thresholded using a universal threshold approach [44]. The original EMG signals and the denoised EMG signal of length 2000 samples are shown in Figure 11 using different thresholding approaches.

4.2. Signal Segmentation and Decomposition by Autocepstrum Analysis

A signal’s segmentation function is to split it up into many epochs with identical statistical properties, like frequency and amplitude [45]. The pre-processing stage for non-stationary signal analysis is typically signal segmentation because stationary signals are easier to analyze than non-stationary signals. Signal segmentation can be divided into two types: adaptive segmentation and constant segmentation. Signals are separated into fixed epochs in continuous segmentation. Constant segmentation has low reliability even though it is straightforward and easy to execute.

Before the feature extraction, each sEMG sample is segmented into 200 ms windows with 20 ms overlap between segments. Segmentation separates muscle contraction from rest, shortening response time without accuracy loss. Some segments contain mostly noise, providing little information. Others show clear patterns during exertion. A few long, convoluted segments displayed alternating contraction and relaxation over several cycles. Overall, the technique revealed the key characteristics of muscle activity in a format suited for further processing while preserving temporal data essential for modeling dynamic movement. The autocepstrum of a given signal essentially refers to the inverse Fourier transform of the natural logarithm of the signal’s power spectral density and provides useful information for signal detection. Computing the autocepstrum of a discrete-time signal involves taking the IFT of the log of its Power Spectral Density (PSD), which is defined by the following:

$$c_a\left(\widehat{n}\right)={\displaystyle \frac{1}{\sqrt{N_r}}}{\Sigma }_{k=0}^{N_r-1}Z\left(\widehat{k}\right)exp\left({\displaystyle \frac{2\pi \widehat{k}\widehat{n}}{N_r}}\right),$$

(2)

where $Z\left(\widehat{k}\right)$ denotes the natural logarithm of the signal’s PSD, $N_r$ denotes the size of the raw sensory signal, $\widehat{n}$ is the quefrency parameter, and $\widehat{k}$ is the discrete frequency parameter. Previously, employing the autocepstrum approach has proven effective in detecting communications signals in low signal-to-noise ratio environments. Research has shown that the autocepstrum of certain signal patterns exhibits a major peak correlating to the reciprocal of the operating frequency along with other minor peaks at multiples of this frequency [46].

On the other hand, there is only one prominent peak at the zeroth quefrency value and very few other peaks in the autocepstrum of additive white Gaussian noise. When analyzing a signal in the cepstral domain, this unique feature is helpful for reducing noise and interference. Figure 12a shows an IMU signal representing ramp-walking activity after being analyzed by the autocepstrum approach, revealing a high peak during the burst interval. Accordingly, our proposed feature extraction method is formulated dependent on this distinguishing observation. To the best of our knowledge, the exploration of raw sensory signals within the autocepstrum domain remains relatively novel and has yet to be substantially discussed in existing literature. The autocepstral peaks can be utilized to distinguish between different activities, considering the peaks’ widths and energy. Figure 12b illustrates the segmented data after rectification and denoising for ramp and walking upstairs.

4.3. Feature Extraction

Feature extraction is the process of identifying distinctive characteristics in segmented data. By eliminating extraneous noise and highlighting significant elements, feature extraction turns unprocessed signal data into a meaningful representation. A sliding window technique is used to accomplish segmentation.

There are several methods for identifying important features, including looking at data in the frequency and temporal domains. This method of signal evaluation facilitates the detection of muscle activity. These characteristics facilitate the classification of unclear recordings with poor signal clarity into structured categories. To take advantage of the autocepstrum’s built-in noise cancellation, we utilize features that have been analyzed via quefrency analysis. Sometimes, to fully extract qualities from highly variable data—where traits are distributed unevenly across sections, further segmentation is required. After that, several statistical features can be extracted, such as the variance, skewness, kurtosis, mobility, complexity and average of autocepstral peaks [47]. The most significant features that affected the results obtained are summarized in

Table 1. Mathematical expressions of the extracted features.

Extracted Feature	Mathematical Expression
Autocepstral Peak	$\mathit{max}{c}_a\left(\widehat{n}\right)={\displaystyle \frac{1}{\sqrt{N_r}}}{\displaystyle \sum _{k=0}^{N_r-1}}Z\left(\widehat{k}\right)exp\left({\displaystyle \frac{2\pi \widehat{k}\widehat{n}}{N_r}}\right)$
Skewness	${\displaystyle \frac{E{(x-\mu )}^3}{{\sigma }^3}}$
Kurtosis	${\displaystyle \frac{E{(x-\mu )}^4}{{\sigma }^4}}$

.

Since this work does not include multimodal signals or sensor fusion, our objective is to decide on the most reliable sensor between sEMG and IMU based on the classification results obtained. In the feature extraction phase, we chose three features, namely, autocepstral peak, skewness and kurtosis of the denoising signals, and we evaluated the average of each feature through all segments. As compared to the work presented in [21], the authors suggested the use of eight handcrafted features for the classification of different human activities, whereas in the proposed manuscript, we utilized only three features to achieve higher classification accuracy.

Therefore, our general preprocessing approach that is explained in Section 4 can be compared to the signal conditioning process approach in [27]. This approach uses combined wavelet denoising and EEMD to investigate its capability for pattern recognition of human activities. The signals of these muscles were preprocessed using wavelet denoising and EEMD. To denoise the IMFs, hard thresholding was adopted, and denoised signals were segmented with 1024 samples per segment with 50% overlapping windowing. The utilized features are slope sign changes (SSC), root mean square (RMS), and average standard deviation value (ASDV) [48]. Classification is based on SVM to classify between different activities using the Georgia Tech. dataset and is conducted for the following three cases:

• Case 1: three activities, namely, slow, fast, and normal walking patterns.

  ○

  Accuracy obtained: 87.5%.

• Case 2: four activities representing walking on stairs with four different heights: ‘Step Height: 4 in’, Step Height: 5 in’, Step Height: 6 in’, Step Height: 6 in’, and ‘Step Height: 7 in’.

  ○

  Accuracy obtained: 89%.

• Case 3: two activities: upstairs and downstairs.

  ○

  Accuracy obtained: 93.5%.

These results, when compared to the ACA-based approach, show that our approach had a better effect on the later classification stage and superior classification results. The details for data classification are shown in the next section.

5. Data Classification, Results and Discussion

The last step in the proposed framework is data classification for activity recognition. The utilized hardware is a computer with an Intel i7 core processor, 2.6 GHz speed and 16 GB RAM. The utilized benchmark dataset, ‘Georgia Tech dataset’ [27], consists of three main activities (walking, stairs and ramp) with a total of 2511 samples collected from 22 able-bodied adults for multiple locomotion modes. The sensors are applied over 11 muscles using EMG sensors sampled at 1000 Hz. Inertial measurement unit data are collected from four different muscles: trunk, thigh, shank, and foot segments sampled at 200 Hz. The samples are divided into training and testing samples for different activities, as can be shown in

Table 2. The details of sensor samples that were used in classification.

Sensor	Activity	Total Number of Samples	No. of Training Samples	No. of Testing Samples
EMG	Walking	199	139	60
	Ramp	1394	952	419
	Stairs	918	642	276
IMU	Walking	221	155	66
	Ramp	1360	952	408
	Stairs	916	641	275

. The proportion of training samples to testing samples is 70% to 30%.

The applied classifiers are K-nearest neighbor, neural networks and random forest. For KNN, the number of neighbors is three, equal distance weighting is used and the distance metric technique is Euclidean distance. The utilized neural network is a feedforward NN with 1 hidden layer, ReLU is the activation function, 100 epochs and stochastic gradient descent as an optimizer. The third classifier, random forest, has several trees equal to seven. For results analysis and assessment, precision (P), recall (R), and F-measure are calculated. Their equations are shown in Equations (3)–(5) as follows:

$$Precision\left(P\right)={\displaystyle \frac{TP}{TP+FP}},$$

(3)

$$Recall(R)={\displaystyle \frac{TP}{TP+FN}},$$

(4)

$$F-measure(F)=2{\displaystyle \frac{P\times R}{P+R}}$$

(5)

where TP represents the number of correctly detected activities, FN represents the number of undetected activities, and FP represents the number of incorrectly detected activities. The full obtained results are shown for EMGs and IMUs in the Appendix A (

Table A1. The complete classification results of KNN, NN and RF on sEMG sensors.

Leg Muscle	KNN %			Neural Network, %			Random Forest, %
	P	R	F	P	R	F	P	R	F
Gastrocnemius	93.9	67.8	78.8	97.8	92	94.8	95.57	96.27	95.92
Tibialis anterior	86.6	60.9	71.5	87.4	82.3	84.8	96.52	90.32	93.31
Soleus	86.7	61.2	71.7	95.2	93.7	94.5	96.95	97.38	97.16
Vastus medialis	84.3	58.3	69	79.5	73	76.1	98.83	96.03	97.41
Vastus lateralis	91.5	65.1	76.1	96.8	94.2	95.5	99.01	97.14	98.07
Rectus femoris	93.2	67.2	78.1	97.7	92.6	95.1	97.88	96.98	97.43
Biceps femoris	79.6	54.5	64.7	81.9	79.1	80.4	98.04	94.84	96.41
Semitendinosus	91.5	67.8	77.9	97.2	92.6	94.8	97.94	94.29	96.08
Gracilis	89	64	74.4	94.9	94.5	94.7	99.53	96.67	98.08
Gluteus medius	88.4	63.2	73.7	97	93.3	95.1	97.66	95.87	96.76
Right external oblique	90	62.7	73.9	92.6	87.4	89.9	96.70	96.27	96.48

and

Table A2. The complete classification results of KNN, NN and RF on IMU sensors.

Sensor Place on the Leg and Its Axis	KNN			Neural Network			Random Forest
	P	R	F	P	R	F	P	R	F
foot_Accel_X	86	68.8	76.5	95.4	91	93.2	99.50	99.92	99.71
foot_Accel_Y	74.7	50.3	60.1	85.3	80.5	82.8	99.12	94.44	96.73
foot_Accel_Z	82.1	60.9	70	96.4	93.4	94.9	98.44	96.81	97.62
foot_Gyro_X	81.1	53.5	64.5	85.5	78.4	81.8	98.44	96.81	97.62
foot_Gyro_Y	82.7	62.5	71.2	94.8	90.9	92.8	98.97	93.43	96.12
foot_Gyro_Z	86.2	71.4	78.1	94.5	88.6	91.5	98.63	97.82	98.22
shank_Accel_X	81.1	64.2	71.7	90.2	85.7	87.9	99.68	97.98	98.82
shank_Accel_Y	81.4	59.6	68.8	92.9	89.3	91.1	99.02	97.39	98.20
shank_Accel_Z	84.8	62.2	71.8	97.8	94.2	96	99.76	98.48	99.12
shank_Gyro_X	93.8	66	77.5	97	93.3	95.1	98.54	97.31	97.92
shank_Gyro_Y	89.6	70	78.7	96.6	91.9	94.2	99.68	97.98	98.82
shank_Gyro_Z	85.5	68.3	75.9	93.6	89.7	91.6	97.65	97.65	97.65
thigh_Accel_X	77.2	52.4	62.4	86.9	80.8	83.7	97.38	97.07	97.22
thigh_Accel_Y	87.8	65.7	75.1	89.7	85	87.2	99.44	96.46	97.93
thigh_Accel_Z	82.2	60.9	69.9	89.6	85.2	87.3	91.41	93.56	92.47
thigh_Gyro_X	91.5	66.6	77.1	89.9	94.4	92.1	90.94	96.95	93.85
thigh_Gyro_Y	83.8	61.7	71.1	90	87.3	88.7	94.60	93.20	93.90
thigh_Gyro_Z	81.5	55.8	66.3	85.2	78.7	81.8	93.04	95.16	94.09
trunk_Accel_X	84.5	59.3	69.7	93.6	84.4	88.8	96.46	94.54	95.49
trunk_Accel_Y	93.7	68.4	79.1	98.8	95.5	97.15	97.50	94.70	96.08
trunk_Accel_Z	86.8	62.8	72.9	94.7	93.2	93.9	93.00	96.25	94.60
trunk_Gyro_X	82.1	56.5	66.9	94.6	91.9	93.2	89.14	94.76	91.87
trunk_Gyro_Y	84.8	60.9	70.9	95.7	92.5	94.1	94.42	93.71	94.06
trunk_Gyro_Z	84.9	59.6	70	95.6	92.1	93.8	96.58	95.05	95.81

), respectively.

For further validation of the obtained results, a threefold K-fold cross-validation technique is used. The dataset is split up into three folds. Each iteration uses a single fold as testing data and the other folds as training data. Consequently, the procedure is repeated until every dataset has been assessed. The mean score of the evaluation metrics values is typically used to represent the K-fold results. The classification results obtained for EMGs and IMUs are shown in

Table 3. Classification results of EMGs after the K-fold validation step.

Leg Muscle	KNN (%)			Neural Network (%)			Random Forest (%)
	P	R	F	P	R	F	P	R	F
Soleus	84.42	59.27	69.64	90.39	86.57	88.44	98.56	96.72	97.63
Vastusmedialis	84.67	59.09	69.59	96.43	90.71	93.48	97.65	95.69	96.66
Vastuslateralis	91.30	65.33	76.16	96.68	92.55	94.57	96.98	95.9	96.41
Rectusfemoris	91.30	65.67	76.39	96.56	91.21	93.80	95.4	94.84	95.0
Gracilis	89.71	64.60	75.11	96.89	92.60	94.68	97.49	96.75	97.11

and

Table 4. Classification results for IMUs after the K-fold validation step.

Sensor Place on the Leg and Its Axis	KNN (%)			Neural Network (%)			Random Forest (%)
	P	R	F	P	R	F	P	R	F
foot_Accel_X	83.43	66.02	73.71	95.00	92.64	93.79	95.61	95.00	95.26
foot_Gyro_Z	86.50	71.67	78.39	91.95	88.46	90.16	97.22	98.05	97.63
shank_Accel_X	82.75	65.18	72.83	90.62	88.29	89.43	95.03	96.28	95.55
shank_Accel_Y	79.70	60.51	68.66	92.19	87.80	89.94	95.14	94.98	94.98
shank_Accel_Z	79.81	61.28	69.27	91.02	71.42	77.45	96.12	97.19	96.61
shank_Gyro_Y	89.14	70.21	78.53	92.63	72.30	78.68	99.29	97.76	98.52

, respectively, where the highest five sensors results in terms of P, R and F are displayed with respect to other sensors.

The results for EMGs in Table 3 show that the random forest algorithm provides higher results than the other techniques. The precision (P) varies from 95.4 up to 98.56%, the recall (R) varies from 94.84 up to 96.75% and the F-measure (F) varies from 95 up to 97.63%. The highest assessment values were achieved for the soleus muscle, where the F-score value was 97.63%. The soleus muscle produces the best results when employing a random forest classifier. Figure 13a displays the confusion matrix of three activities (walking, ramping, and stairs) using an EMG sensor. According to the results, all samples were accurately recognized for the third activity (stairs); however, a few samples were incorrectly detected for the first and second activities (walking and ramp).

Similarly to EMG, the random forest classifier outperforms the other two classifiers in terms of the aforementioned metrics when utilizing an IMU sensor. The precision (P) varies from 95.03 up to 99.29%, the recall (R) varies from 94.98 up to 98.05% and the F-measure (F) varies from 95.05 up to 98.52%. The best results can be obtained by extracting the gyroscope signal in the Y direction from the shank muscle.

The classification results of the three gyroscope signal activities in the Y direction that were taken from the IMU sensor using a random forest classifier are displayed in the confusion matrix in Figure 13b. The findings showed that while all samples were correctly identified for the third activity (stairs), several samples were misidentified for the first and second activities (walking and ramp). The results are shown in Table 4.

The previous results indicate that the best discrimination between the different activities can be achieved by using an EMG signal collected from the soleus muscle or by using a signal extracted from an IMU sensor located on the shank muscle. These obtained results can be compared to other HAR approaches that are mentioned in the literature, as can be seen in

Table 5. Comparison between our activity recognition approach and other HAR systems.

Name	ADLs	Utilized Dataset	Utilized Classifiers	Utilized Sensors	Accuracy Indicators
New ACA-based framework (our approach)	Walk Stairs up Stairs down Ramp walk	Georgia Tech. (22 subjects)	KNN NN RF	EMGs IMUs	97.63% 98.52%
Tseng et al. [22]	Stand Sit Walk Run	Custom experimental dataset	XGBoost CVAE	IMUs	96.03%
Vijayvargiya et al. [49]	Sitting Standing Walking circuit	EMG dataset for lower limb [50] (22 subjects)	CNN-based classfiers LSTM	EMGs	99.86%
Kudus et al. [51]	Walk Stairs up Stairs down Ramp walk	Georgia Tech. (22 subjects)	KNN SVM-based classifiers	IMUs GONs	90.1%

. Hence, the results achieved can be seen as comparable and successful to other research results.

6. Conclusions and Future Work

In this paper, we have proposed an activity recognition framework based on signal segmentation, decomposition, and feature extraction from sEMG and IMU sensors. This approach is applied to two different types of signals extracted from EMG and IMU sensors. This methodology relies heavily on deep signal conditioning of IMU and EMG signals to pave the way for easily implemented machine learning classifiers. Autocepstrum analysis (ACA) was chosen for signal conditioning after several trials with other techniques like WT, EMD and PCA. Three machine learning classifiers were chosen, and they were able to achieve superior accuracy for locomotion activities, particularly in gait rehabilitation applications. The resultant data were assessed by confusion matrices, precision, recall and F-measure indicators. The data were later validated by the K-fold validation technique.

The results indicate that the random forest classifier performs better than KNN and neural networks across all muscle groups. As for EMG signals, the most accurate results were obtained from the soleus, gracilis, and vastus medialis muscles, with F-scores of 97.63%, 97.11% and 96.66%, respectively. On the other hand, shank and foot signals achieved the highest the F-scores, with 98.52% and 97.63%, respectively. These findings validate the necessity of sensor-based HAR for rehabilitation robotic devices.

Future work will focus on obtaining a new custom dataset for real-time data collection and deep learning-based classification. Further research will include the development of a new data acquisition system with the previously recommended classification framework to be integrated with an actual rehabilitation device to enhance mobility assistance for individuals with neuromuscular impairments. Indeed, high-performance sensors with advanced specifications, such as precise calibration, reliable data transmission, and seamless integration with lower limb rehabilitation devices, must be carefully considered to ensure the accuracy and practicality of sensory data acquisition in real-world applications. Moreover, clinical validation of the acquired data with future rehabilitation devices should be monitored carefully to ensure its effectiveness, safety, and applicability in real-world therapeutic scenarios, particularly for patients with varying levels of mobility impairments. Finally, further investigation of using multimodal sensory data for signal processing and classification is planned.