Forearm Radiomyography for Character and Writer Identification in Air Writing

Abstract

This paper explores the air writing application of radiomyography (RMG), a near-field radio frequency (NFRF) sensing modality, and studies the key issues in human-computer interaction (HCI), including user acceptability and privacy concerns. Our primary objective is to demonstrate the efficacy of our NFRF sensing system in identifying both handwritten characters and the writer. We employed a forearm-worn multiple-input-multiple-output (MIMO) RMG to capture the detailed actuation of superficial and deep muscles during the air writing motion through enhanced electromagnetic coupling in the near-field. The sensing architecture leverages synchronized software-defined radios (SDRs), custom antenna arrays, and frequency-division multiplexing to enable precision muscle activity detection in a wearable armband. Various signal processing and machine learning (ML) algorithms were benchmarked for both classification accuracy and computational efficiency. The experimental study included 26 letters and 25 writers. The best performing algorithm achieved an average character classification accuracy of 92.35% (95% CI: 90.06–94.64). Moreover, through transfer learning, our model reached 95% testing accuracy in character classification for new subjects using just 10% of the available data for training. The writer identification model achieves an average accuracy of 99.1%, including longitudinal studies. These results highlight the promise of forearm-worn RMG as a novel HCI alternative, offering accurate performance without the need for cumbersome devices on the hand that may obstruct dexterity or the need for wet body-contact electrodes in surface electromyography.

I. Introduction

Human-computer interfaces (HCI) have evolved over the past 25 years from keyboards and mice to more intuitive touch screens. With growing demand to reduce physical interfaces, alternatives like voice recognition encounter challenges such as ambient noise and eavesdropping issues [1]. Mid-air gestures offer promise [2] but are constrained by predefined movements that restrict interaction possibilities and require user memorization [3]. Air writing, maneuvering complex character writing with subtle hand movement in the air, addresses these issues by enabling rich and familiar input without learning, and remembers new gestures. Air writing is also a more practical and precise alternative in environments prone to noise or requiring high privacy [4].

Air writing involves forming letters in free space using a pen-like device or bare fingers [1]. Unlike traditional writing, it lacks haptic and visual feedback, as well as motion of pen ups and downs, making it more complex and uncertain for character recognition [1], [5]. Still, air writing provides a useful alternative to typing or touch pad [6], especially for covert communication where privacy or hand freedom is of critical importance [7]. In comparison, current systems using vision or radar with off-body readers face issues like privacy, line-of-sight (LoS) blockage, cost, and off-body setup complexity [4], [8]. Wearable accelerometer-based methods on hands and fingers [9] may also reduce dexterity, even with glove integration. Surface electromyography (sEMG) requires stable skin-contact electrodes and conductive interfaces, which may introduce variability due to skin impedance changes and can limit wearability during extended use [10].

Many hand movements are derived from forearm muscles via tendons, enabling strength without impairing fine finger control from muscular weight on palms and fingers. This makes forearm muscle sensing a practical and user-friendly interface for air writing. Individual differences in muscle activity suggest that forearm sensing could possibly reveal both written content and biometric traits. These differences can support free-space signatures for identity verification, which are similar to air writing for character recognition but put primary emphasis on biometric traits [11]–[17]. Like air writing, air signature lacks visual and haptic feedback, and thus is more challenging in terms of reliability and repeatability. While current systems rely on cameras and accelerometers, internal muscle activity is more distinctive and harder to replicate [14], offering a more secure option for identity verification. If only character recognition is desirable, understanding of the biometric content in the recording can also help de-identify the personal characteristics.

First introduced in [18], RMG enables continuous, noninvasive monitoring of superficial and deep muscle contractions using the near-field coherent sensing (NCS) principle [19] by wearable near-field radio-frequency (NFRF) transceivers and sensing elements. RMG supports multiple-input-multiple-output (MIMO) setups for rich observation diversity and requires no direct skin contact. Wearable RMG is LoS-independent, resilient to obstructions, and allows free user motion indoors and outdoors. Unlike surface-based methods such as imaging and sEMG, RMG captures internal tissue dynamics, as shown in tracking heart valves [20], diaphragmatic motion [21], and vascular pulses [22]. These features make RMG a strong candidate for air writing and air signature via forearm muscle sensing.

A. Related Works

Air writing and air signatures have been considered as important methods within HCI in the literature. This section reviews relevant research on these fields.

Air writing has been explored in the systems of cameras, radars, inertia-measuring units (IMU), and sEMG. Imaging and radar systems are the main off-body techniques. Vision-based systems track 2D/3D motion using webcams [8], [23], colored markers [24], or depth sensors [4], [25]. Radar systems commonly use millimeter-wave devices such as networks of 60 GHz continuous-wave (CW) coherent transceivers [6], single ultra-wide-band (UWB) pulse radars [26], or disturbance of ambient Wi-Fi [27]. Despite their versatility, off-body sensors face challenges including fisheye distortion, LoS blockage, poor low-light performance [24], limited capture volume, ambient interference, and high system complexity from multiple observation locations to relieve LoS blockage and enhance depth perception.

Wearable air writing systems have explored IMUs on the fingers [9], wrist [28], or hand [1], [3], [5], [29], [30]. However, wrist-worn systems can be ambiguous, and finger-worn or handheld devices are often bulky, affecting dexterity. Alternatives like sEMG and RMG offer promise for smartwatch or forearm-band designs, which provide greater comfort and convenience. As a proof of concept, sEMG in [10] demonstrated air writing recognition using a forearm band.

Though less studied than air writing, air signature has gained interest in recent works, mostly using vision-based systems. Some studies used video cameras [11], [12], depth sensors [15], or Leap Motion^™ controllers [17], [31]. These methods, however, are vulnerable to presentation attacks using robotic arms and hidden cameras. Beyond vision, sensing methods had also been expanded to millimeter-wave radars [13] and IMUs [16], similar to the studies in air writing.

In summary, each air writing and signature method has trade-offs in performance and convenience. Radar systems are robust but complex to implement in multi-location setups and sensitive to interference. Vision-based methods offer accurate tracking but struggle with occlusions and lighting. Wearables like IMUs are close to the motion source but face ambiguity and discomfort. sEMG offers direct muscle sensing and wearable integration, but it is less accurate and sensitive to body electrode contact quality.

B. Major Contributions

In this paper, we demonstrate the applicability of MIMO-based RMG for air writing and writer identification, extending its prior use in hand gesture recognition [18]. We collected continuous forearm muscle actuation transients from 25 participants drawing all 26 English upper-case letters in the air. The resulting 1-D time-domain waveforms were filtered and transformed into 2-D time-frequency spectrograms using continuous wavelet transform (CWT), then classified using a conventional convolutional neural network (CNN). We benchmarked multiple signal processing and ML algorithms, achieving high accuracy in both character and writer recognition. This work presents the first feasibility study of using direct monitoring of forearm muscle activities in air writing and biometrics, which offers new alternatives for privacy, convenience, and comfort in HCI applications.

II. RMG Principles

NCS [19] is an NFRF technique that detects the movement of dielectric boundaries by leveraging effective electromagnetic (EM) coupling in the near field of the antenna. RMG builds on the NCS principle to monitor human skeletomuscular motion in a continuous and contactless manner [18]. This approach avoids many limitations of contact-based sensors, making it suitable for long-term monitoring of muscle dynamics. RMG operates in the ultra high frequency (UHF) band between 300 MHz and 3 GHz [32], which allows for efficient EM energy coupling into the body using wearable antennas. To achieve high spatial diversity of observation, RMG can be implemented with a MIMO configuration, which enables simultaneous sensing at multiple locations.

Each RMG sensing channel consists of a transmitter (Tx) that delivers EM energy to the target area and a receiver (Rx) that detects signals modulated by dielectric boundary motion from muscle contraction and permittivity changes from blood flow. In the forearm, biological tissue actuation alters the antenna impedance and near-field scattering properties. These changes are captured as variations in the Rx signal and recovered through carrier demodulation and baseband filtering [19]. The Rx signal contains two components: the magnitude reflects motion relative to the antenna, while the phase captures motion that occurs synchronously with the antenna [19]. To capture the spatial complexity of muscle contractions, RMG uses a MIMO architecture in which N sensing points produce N² independent Tx-Rx combinations. Due to tissue dispersion and near-field nonlinearity, each Tx-Rx pair encodes unique physiological information, including the reciprocal pairs. Additionally, from the perspective of conventional microwave circuits, the MIMO output can also be represented as a multi-port scattering (S) matrix. Channel isolation is achieved using frequency division multiple access (FDMA), which allows more than −60 dB of cross-channel isolation [18].

RMG offers several advantages over conventional wearable methods, such as sEMG [33] and electrical impedance tomography (EIT) [34], which rely on skin contact and are vulnerable to noise from motion artifacts and electrode instability. RMG does not require direct electrical contact to the skin or adhesive electrodes, as sensing is based on near-field electromagnetic coupling rather than quasi-static Ohmic interface. This also allows the sensors to be worn over clothing. While far-field radar systems typically detect only surface-level motion due to limited tissue penetration, NCS enables deep tissue sensing by confining the interaction to the near field of the antenna. As a result, RMG achieves higher spatial resolution and localized sensing. Moreover, it operates independently of lighting conditions, requires no LoS, and promotes user privacy without camera imaging. Its implementation on mature Radio Frequency (RF) platforms also makes it a scalable and cost-effective solution.

The forearm consists of anterior and posterior compartments, comprising 20 muscles that control elbow, wrist, and finger motion [35]. Surface muscles are primarily involved in wrist motion, while deeper muscles control fine finger articulation. In our study, participants wrote characters using finger and mild wrist motion with the elbow anchored. Table I summarizes the forearm muscle groups captured in this study, and Table II maps them to wrist and finger actions relevant to air writing. Thus, this application requires monitoring of muscle motion from all compartments of different depths. To accommodate this complexity, the RMG system in this study used a 4 × 4 MIMO array, resulting in 16 independent observation channels. This configuration enabled well-covered mapping of forearm muscle activities during air writing.

TABLE I.

Forearm muscle anatomy

Compartment		Muscle	Acronym
Anterior	Superficial	Flexor Carpi Ulnaris	FCU
		Palmaris Longus	PL
		Flexor Carpi Radialis	FCR
		Pronator Teres	PT
	Intermediate	Flexor Digitorum Superficialis	FDS
	Deep	Flexor Digitorum Profundus	FDP
		Flexor Pollicis Longus	FPL
		Pronator Quadratus	PQ
Posterior	Superficial	Extensor Carpi Radialis Longus and Brevis	ECRL/B
		Extensor Digitorum Communis	EDC
		Extensor Digiti Minimi	EDM
		Extensor Carpi Ulnaris	ECU
		Anconeus	A
	Deep	Supinator	S
		Abductor Pollicis Longus	APL
		Extensor Pollicis Brevis	EPB
		Extensor Pollicis Longus	EPL
		Extensor Indicis Proprius	EIP

TABLE II.

Associated muscle groups for airwriting

Joint	Motion type	Associated muscles
Wrist	Flex	FCU, PL, FCR, FDS
	Extend	ECRL/B, EDM, ECU
	Ulnar Deviation	FCU, ECU
	Radial Deviation	FCR
Index and middle fingers	Flex	FDS, FDP
	Extend	EDC, EIP

III. Methods

A. Experimental setup

1). RMG implementation:

The forearm was selected as the sensing site due to its critical role in initiating wrist and finger movements, its high signal entropy, and its suitability for comfortable wearing. Both left- and right-handed participants were included in the study, with antenna placements adjusted accordingly. In our preliminary prototype for a feasibility study, a flexible textile armband prototype was used to secure the sensing system around the forearm. Improvement on packaging and user comfort will be performed in the future product development phase. Each sensing point included two whip antennas (Taoglas TG.19.0112), one Tx and one Rx, mounted on 3D-printed holders and affixed using Velcro straps to allow adjustable and secure placement. The antennas were aligned parallel to the muscle fibers and distributed to cover relevant forearm muscle groups, as detailed in Table I, with placement shown in Fig. 1(a).

Fig. 1.

Experimental setup. (a) A diagram of the placement of the RMG antennas and sEMG electrodes; (b) Schematic diagram of signal connection.

The whip antennas used at each sensing point are electrically short monopoles that exhibit a quasi-omnidirectional radiation pattern in the azimuth plane under free-space conditions. In the present configuration, however, the antennas operate in close proximity to biological tissues, where electromagnetic interaction is dominated by localized near-field coupling rather than conventional far-field radiation behavior. When mounted directly on the forearm, tissue loading alters the effective radiation characteristics, resulting in spatially confined electromagnetic coupling to the underlying muscle structures in the near field [18]. As a result, classical far-field side-lobe behavior plays a limited role in signal formation and environmental clutters are expected to be weak. Potential interference from mutual MIMO channel coupling is mitigated through frequency multiplexing. These measures ensure that the received signals predominantly reflect local dielectric modulation associated with muscle contraction.

The system employed two Ettus B210 software defined radios (SDRs), enabling a 4 × 4 MIMO configuration to support 16 independent sensing channels. Each SDR supported two synchronous Tx-Rx pairs and was connected to a host computer via universal serial bus (USB). The SDRs were synchronized using an external local oscillator (BG7TBL-GPSDO), which provided a 10 MHz reference clock and a 1 pulse-per-second (PPS) baseband timing signal. Control software was implemented using LabVIEW. Antennas on the armband were connected to the off-body SDRs via subminiature version A (SMA) cables. This implementation can be observed in Fig. 1(b).

Each Tx generated a digital baseband tone that was converted to the analog form via a digital-to-analog converter (DAC), which was then mixed with the carrier frequency using a quadrature superheterodyne scheme. The carrier frequency was set at 900 MHz to achieve feasible near-field penetration, given that the in-body wavelength at this frequency (5 cm) matches the typical forearm diameter, assuming a relative dielectric constant of 36 – 64 [18]. Tx power was kept below −10 dBm (0.1 mW), fully compliant with Occupational Safety and Health Administration (OSHA) guidelines for UHF transmission [36]. An FDMA scheme was used to isolate the four Tx, each operating at different intermediate frequencies of 10, 25, 40, and 125 kHz, respectively. The two SDRs were configured to capture 4 self-backscattering and 12 cross-channel measurements, resulting in a total of 16 MIMO channels. Raw signals were sampled at 1 million samples per second (sps) and subsequently down-sampled to 500 sps for further processing.

2). sEMG implementation:

sEMG has been broadly studied in gesture recognition and air writing for HCI [29], [37]–[40]. In selected sessions, we collected synchronous sEMG data to compare with RMG. Two sEMG channels were recorded using a commercial device (BIOPAC MP36R). This configuration does not represent a high-density or optimized sEMG system but was selected to provide a synchronized reference under identical preprocessing and classification pipelines. Optimized multi-electrode sEMG can have much better performance than our benchmark results here. Nevertheless, this comparison helps validate our recognition algorithms and highlights different physiological features in RMG and sEMG. The BIOPAC system used SS2LB leads and EL503 column electrodes. Each sEMG channel included positive and negative electrodes, and the ground electrode was shared. Channels were placed on the anterior and posterior sides of the forearm, with the ground electrode on the wrist, which had mostly tendons and ligaments with the least muscles to provide a stable sEMG reference. The signal electrodes on either side of the RMG armband were shown in Fig. 1(a). Electrode placement was adjusted based on the signal strength in hand gesture responses [18], and LabVIEW was used to synchronize RMG and sEMG recordings.

3). Human study protocols:

This study was approved by Cornell IRB Protocol #1812008488 and conducted on 25 healthy participants with written informed consent. Each subject wrote all 26 uppercase letters in the air using their index and middle fingers, assisted by wrist motion, while the elbow remained supported and stationary (Fig. 1(a)).

This setup addresses the specific challenges of air writing, which lacks haptic and visual feedback and involves an unconstrained, imaginary surface. Since our focus is on evaluating RMG applicability to individual letter recognition, not word-level analysis, we limited the task to the 26 uppercase English letters. Participants wrote each character within a self-imaginary box. Early trials using only the index finger produced weaker muscle signals and lower accuracy, so both index and middle fingers were used. To aid signal segmentation, a brief resting period was introduced between letters. This approach enabled simple data segmentation, deferring more advanced timing methods for future work.

To build a robust machine learning (ML) dataset, each participant performed 20–25 repetitions of all 26 uppercase letters, balancing data volume with collection time. The letters were divided into two routines, each of which was around 4 minutes and contained a mix of similar and distinct characters. A fixed 5-second epoch with voice prompts was used for each letter. Repeated routines and scheduled breaks helped to reduce fatigue and bias. As shown in Fig. 2, distinct waveforms were observed across most MIMO channels, suggesting that each captured different muscle information.

Fig. 2.

Sample waveforms for two subjects and the letters “N” and “O” in different RMG channels and one sEMG channel.

In free-space air writing, the same character can be produced using different stroke orders or directions. While these variations may result in similar external trajectories, they can lead to substantially different patterns of forearm muscle activation. This effect is particularly relevant for RMG and sEMG [10]] sensing, which measures intrinsic muscle activities rather than external motion. As a result, variations in stroke execution can significantly alter the temporal pattern of muscle recruitment for the same character, effectively increasing intra-class physiological variability, as different subjects may execute the same character using different stroke orders or directions. This challenge is further exacerbated for character pairs such as (C, O), (D, P), and (M, W), where similar writing motions can produce nearly indistinguishable muscle activation patterns, reducing separability in the measured RMG and sEMG signals under unconstrained execution. Since the objective of this study was to classify the complete character set rather than reconstruct individual strokes, fixed stroke sequences were defined for each letter (Fig. 3). Particular attention was given to commonly confused character pairs such as (C, O), (D, P), and (M, W), motivated both by prior air writing studies [30] and by intuitive observations made during the initial phase of protocol development. Participants were instructed to follow predefined motion directions (e.g., drawing “C” anticlockwise and “O” clockwise) to ensure consistent execution across characters and subjects.

Fig. 3.

Fixed stroke sequences for the capital alphabets in two routines. The ambiguous letter pairs are shown in colored boxes where specific stroke sequences are introduced to improve differentiation.

Potential confounding factors in forearm muscle sensing include differences in age and body composition, variations in individual motor strategy, writing speed, fatigue, and minor changes in sensor placement. As this work is intended as a feasibility study of RMG for air writing, the data collection was conducted under controlled lab conditions to better understand the intrinsic sensing capability of the system. The experimental posture and writing protocol were standardized for all participants, and predefined stroke sequences were used to reduce behavioral variability. Variability in writing speed was addressed during signal processing using dynamic time warping (DTW). Antennas were positioned according to anatomical landmarks and secured to maintain stable coupling throughout data collection. During the rest period between the 4-min routines, the sensor position was slightly adjusted by the operator or subject for comfort. The position variation was included as part of the data variability considered in the classification algorithm. These measures were designed to ensure that the recorded signals primarily reflect muscle activation patterns associated with character formation and writer identity, rather than unrelated sources of variability within the studied cohort.

In addition to controlling experimental factors, motion conditions were selected to emphasize intrinsic forearm muscle dynamics. The current prototype employs commercial omnidirectional antennas designed for general-purpose RF operation rather than spatially selective muscle sensing. Under this configuration, large-scale arm motion introduces additional motion interference that may obscure fine-grained muscle activation patterns. Accordingly, experimental posture and gross arm motion were minimized to evaluate baseline RMG sensitivity to localized muscle activities.

The dataset collected from this study will be made available upon reasonable request.

B. Signal Processing

1). Preprocessing:

The same preprocessing pipeline was used for both character classification and subject identification. First, a 0.01–10 Hz bandpass filter was applied to remove high-frequency noise and retain target motion signals. Waveforms were then normalized by centering at zero and scaling to unit standard deviation for consistency. Detrending followed, removing linear drift within each epoch. To account for variations in writing speed, DTW was applied. Each segment was annotated using voice prompts, and the writer’s identity was recorded by subject case number to ensure privacy.

2). Data segmentation:

Two segmentation methods were evaluated. The first used fixed five-second windows based on voice prompts, capturing both writing and resting periods for consistency, as shown in Fig. 4(b). While fixed segmentation was effective for controlled settings, it would hardly be suitable for practical scenarios. The second approach is the adaptive segmentation, which removes the resting periods by estimating the interval of high activities. Using the Tx1-Rx2 channel as a timing reference due to its high signal-to-noise ratio (SNR), as shown in Fig. 4(a), we computed the moving standard deviation (stdev) of signal amplitude to detect onset and offset time points. To avoid errors from mid-letter fluctuations, multiple peaks in standard deviation were identified, and the outermost ones were used as the crop points. Although this may retain some rest-period data, it ensures minimal loss of informative signals. The identified boundaries were applied uniformly across all channels, and the resulting segments were resampled to a fixed length as shown in Fig. 4(c).

Fig. 4.

Data pre-processing by segmentation. (a) A sample time-domain waveform with moving stdev and the cropping points; (b) Data using fixed segmentation for the alphabets of A, B and C in the amplitudes of Tx2-Rx2, Tx1-Rx2 and Tx3-Rx3; (c) Corresponding data using adaptive segmentation.

3). Warping time series:

DTW is a well-established method for comparing and aligning time series [41]. Widely used in speech recognition [41]–[43], DTW is particularly effective for handling time deformations and varying speeds, making it well suited to air writing, where characters may be written at different speeds and with uncertain start times. We adopted the DTW algorithm from [43] to identify a representative waveform for each letter in each channel. All waveforms were then aligned to this template by minimizing the DTW distance. However, since our goal includes both character and writer identification, aligning data across subjects could introduce bias. Therefore, alignment was performed only within each subject. Additionally, as different channels capture distinct muscle activities, alignment across channels was not meaningful. Thus, alignment was applied separately to each channel within each subject. Examples of DTW processed waveforms are shown in Fig. 5.

Fig. 5.

Sample Tx1Rx2 waveforms for the character “A”: (a) After cropping the beginning and ending points before DTW; (b) Cropped and warped waveforms with the DTW average in the thick blue line.

4). Time-frequency analyses:

Time-frequency analyses allow for richer signal content examination. Short-time Fourier transform (STFT) and CWT are among the most common methods. While STFT extracts instantaneous frequency using a moving window, it involves tradeoffs in time-frequency resolution and lacks adaptability [44]. As observed in Fig. 4, our time series contained both high-frequency transients and low-frequency drifts, which called for a more flexible time-frequency view. CWT, with its variable resolution, proved more suitable and was validated through performance benchmarking. We evaluated five time-frequency methods: two STFT variants with 0.6 s and 1 s windows, and three CWTs using Gaussian (CWT1), Ricker (CWT2), and Morlet (CWT3) wavelets. The Gaussian CWT showed the best performance and was selected for further analyses. Sample time-frequency plots for two waveforms of the alphabet “J” and “M” are shown in Fig. 6.

Fig. 6.

Time-frequency spectrograms for letters “J” (left column) and “M” (right column): (a) Time-domain waveforms; (b) CWT with Gaussian and Ricker wavelets; (c) STFT with 0.6 and 1 second windows.

C. Classification Algorithms

Two ML classification algorithms, vision transformer (ViT) and CNN, were investigated for character classification and subject identification. Even with differences in labeling and data handling, both algorithms share similarities in their main design. We aimed to achieve a balance between computational cost and performance. Given the large number of classes and the relatively high volume of input data, we opted for a less expensive CNN model.

1). Character classification:

For character classification, both personal and global training models were implemented. For global training, subject-specific information was neglected, and the labeling and the follow-on training model were done solely based on the characters written. Notice that the number of our present participants is still small, and many muscle conditions, such as age and body mass index (BMI), were not sufficiently represented. Hence, the global training model yielded low accuracy, which may be significantly improved when the number of subjects becomes much larger or when transfer learning is adopted.

2). Personal identification:

For the writer identification, after a few trials, the best method would need to consider both subject and character information when data were fed into the ML model. Initially, all characters written by each subject were considered for model training. However, the classification accuracies were poor due to high ambiguity in some characters. Consequently, we decided to reduce the number of characters fed into the ML model. To determine which characters contributed the most to the writer classification, we searched different combinations. The four characters that made the most significant differentiation of writer identity were selected, which yielded much more favorable results.

IV. Results and Discussion

A. Air Writing

Fig. 7 depicts the accuracy results obtained from various personal training models for both RMG and sEMG data. Fig. 7(a) shows the average accuracy across different preprocessing methods. It is evident that while fixed segmentation performed around 80% accuracy, the combination of adaptive segmentation and DTW time alignment yielded the best results. Additionally, although the average accuracy across each preprocessing method is comparable, there are noticeable differences in their respective spreading ranges. Notably, CWT1 (Gaussian wavelet) exhibited a smaller deviation, making it the preferred time-frequency conversion method for this specific application. Additionally, the performance of CWT1 is summarized in Table III, which reports mean accuracy with 95% confidence intervals (CI) along with the macro-averaged precision and F1 score.

Fig. 7.

Character classification results averaged over all subjects. (a) The average accuracy of RMG using fixed segmentation (blue), adaptive segmentation (green), and adaptive segmentation + DTW (red) for the preprocessing methods; (b) Accuracy spread of RMG and sEMG data for different preprocessing methods.

TABLE III.

Additional Performance metrics (CWT1)

Modality	Accuracy (95% CI)	Macro Precision	Macro F1
RMG	92.35% (90.06–94.64)	92.61%	92.48%
sEMG	54.34% (43.20–65.48)	54.55%	54.44%

Fig. 8 shows the effectiveness of transfer learning in the global training models. The first model combines data from all 25 subjects into a single dataset, with a subsequent random split irrespective of subject labels, resulting in a low average accuracy of 63.3% in character recognition. This is not surprising, as ignoring the writer’s identity in the data set will cause significant confusion in classification. In contrast, the second model incorporates transfer learning for new subjects. Initially, the model is first trained by leaving a participant out, followed by transfer learning using a given percentage x of the unseen person data, and then testing on the remaining 1 – x data. The percentage x was varied from 0% to 90%. As seen in Fig. 8(a), when the personal training data set is very low, the accuracy starts from around 65%, close to that of the original global training. The standard deviation was evaluated from the rotation of the leave-one-out tests with a full model reset. Average accuracy reaches approximately 95% when only x = 10% was used for training. This shows that personal training for character recognition can be much shorter in practical scenarios when a general model from a large data set is available.

Fig. 8.

Average accuracy for the global training models. (a) The average accuracy of the new participant with different training-testing data splits during transfer learning; (b) The average accuracy of all participants without and with transfer learning (x = 10%). The leave-one-out participant was rotated to generate the distribution.

Furthermore, the comparison between the two ML methods of ViT and CNN is shown in Table IV, where the less expensive CNN often outperformed ViT in terms of character classification accuracy. Notice that this comparison was conducted only on 10 participants, utilizing fixed segmentation and Gaussian CWT. This was due to the fact that the ViT model requires a substantial computational time when the input set becomes large.

TABLE IV.

Comparison of character recognition accuracy between CNN and VIT models

Subject	1	2	3	4	5	6	7	8	9	10	Avg.
CNN (%)	98.2	70.9	90.5	98.4	90.3	91.4	87.7	95.1	91.9	89.7	90.41
ViT (%)	97.4	73.5	81.5	88.6	86.2	88.3	87.4	96.2	91.4	86.4	87.40

B. Writer identification

For writer identification, fixed segmentation without DTW alignment was used to avoid normalization-induced bias, focusing on writer traits rather than character features. Gaussian CWT was selected based on prior experience.

For writer identification, initial CNN training with a global dataset and k-fold cross-validation yielded only 54% accuracy, as character-specific variations masked personal traits. To address this, the dataset was partitioned into letter-specific subsets, and the same algorithm was applied. Each letter then exhibited higher performance. The classification accuracy was highest for the characters “M”,“E”, “W”, and “R”, which involved more complex stroke sequences and broader muscle engagement (Fig. 9(a)). With simple majority voting from the four characters, the model’s overall accuracy improved to 99.1%. The confusion matrix for the final model is shown in Fig. 9(b). Here, the subject index permutation aggregates errors in the upper-left corner, while all remaining subjects achieved perfect identification without false positives or negatives. Subject 1 was misidentified as Subject 23 and Subject 9 a few times, resulting in a true positive rate of 0.95 for Subject 1. Similarly, Subject 23 was misidentified as Subject 1 infrequently, leading to a true positive rate of 0.86. All other subjects achieved a true positive rate of 1.

Fig. 9.

Accuracy for the writer identification. (a) Results from the seven best performing individual letters and the majority voting; (b) Confusion matrix of the majority voting model.

The relatively small dataset enables rapid writer identification, and users can write just a few predefined letters as an identification template. To test robustness, we also collected longitudinal data from five participants on different days and incorporated it into the dataset. These results demonstrate the model’s ability to recognize writer identity across varying sessions and setup adjustments.

V. RMG and sEMG comparison

Synchronous sEMG data were collected during 10 sessions and processed identically to RMG data using adaptive segmentation, DTW alignment, and CNN classification on STFT and CWT spectrograms. As shown in Fig. 7(b), sEMG with two channels achieved approximately 60% accuracy with a larger standard deviation, while MIMO RMG reached around 92%. As further summarized in Table III, the confidence interval for RMG remained relatively narrow, indicating stable performance across folds. In comparison, the wider CI observed for sEMG reflected greater variability under identical preprocessing and classification conditions.

It should be noted again that the two-channel sEMG configuration used here does not reflect the full capability of modern high-density or optimized multi-channel sEMG systems. While additional channels and modality-specific optimization can improve sEMG performance, the present comparison is not intended as a definitive modality-level benchmark. Rather, the inclusion of a deliberately low-dimensional sEMG reference under identical preprocessing and classification conditions serves as a check against potential machine learning overfitting, as overfitting would be expected to yield similarly high performance even with a small number of sEMG channels.

VI. Comparison to previous air writing work

To contextualize the performance of the proposed RMG framework, Table V summarizes representative air writing systems based on vision, IMU, radar, WiFi, and sEMG sensing modalities. All reported results correspond to English capital letter recognition under isolated character classification settings. When both user-dependent and user-independent evaluations were available, the user-dependent performance is presented for consistency, as this protocol is the most commonly reported setting across prior studies. In this work, both evaluation schemes were conducted, and the user dependent results are shown for direct comparison. The listed recognition methods indicate the primary classification approach used for character inference; preceding preprocessing stages are not extensively included in the comparison.

TABLE V.

Comparison with Representative Air-Writing Systems

	Mukherjee 2019 [8]	Zhang 2013 [25]	Arsalan 2019 [6]	Fu 2019 [27]	Arsalan 2020 [45]	Moazen 2016 [30]	Yanay 2020 [1]	Younas 2021 [9]	Tripathi 2023 [10]	This Work
Sensing Modality	Webcam	Kinect	Radar	WiFi	Radar	IMU	IMU	IMU	sEMG	RMG
Dataset Size	1,800	375 videos	–	26,000	3,750	520	21,450	5,270	13,000	13,650
# Classes	26	26	10	26	10	26	26	26	26	26
# Subjects	5	–	–	8	–	1	55	20	50	25
Recognition Method	CNN	MQDF	ConvLSTM	GMM-HMM	1D-TCN	DTW	CNN/DTW	TR	2D-CNN	CNN
Accuracy	96.11%	94.62%	98.33%	88.74%	99.11%	71%	83.20%	95.01%	78.50%	92.35%

Abbreviations: MQDF = modified quadratic discriminant function; ConvLSTM = convolutional long short-term memory; GMM-HMM = gaussian mixture model-hidden markov models; TCN = temporal convolution network; TR = Trajectory Reconstruction

Vision-based approaches [8], [25] reported high recognition accuracies for isolated character tasks but relied on LoS imaging systems and controlled visual environments. Radar-based systems [6], [45] also demonstrated strong performance while requiring off-body RF infrastructure. IMU-based solutions [1], [9], [30] enabled wearable implementations through trajectory reconstruction or motion modeling, whereas sEMG approaches [10] leveraged muscle activation signals but typically required multiple electrodes with stable skin contact, which were difficult for large muscle actuation or long-term recording. Across modalities, dataset sizes ranged from hundreds to tens of thousands of samples, and subject counts varied from single user to above 50, with reported accuracies spanning approximately 70% to 99%. Within this context, the proposed MIMO RMG framework achieved competitive performance using a forearm-wearable configuration based on intrinsic muscle actuation, without reliance on external imaging systems or high-density body electrode arrays. Reported values should be interpreted cautiously, as experimental protocols and data characteristics differed across studies.

VII. Conclusion

Our study demonstrates the potential of MIMO RMG for air writing and air signature, extending its application beyond hand gestures to more diverse HCI methods. The full system prototype achieved high classification accuracy for both air-written characters and writer identity, based on unique muscle motion patterns. Through benchmarking the performance of signal processing and ML algorithms, we underscore the efficacy of forearm-wearable RMG as both an intuitive HCI interface and a dynamic biometric tool during air writing.

While this study was conducted under controlled posture to establish the application feasibility, the findings remain relevant to practical air writing scenarios. Many realistic text-input tasks, such as short command entry, password authentication, and signature verification, are naturally performed under stationary or semi-stationary conditions, where large-scale arm motion would be limited. Within this context, the proposed RMG-based approach demonstrates robust sensitivity to fine-grained forearm muscle dynamics associated with character formation. Moreover, because RMG does not rely on visual LoS, it offers a wearable and camera-free alternative to vision-based air writing systems, making it well-suited for low-light environments, privacy-sensitive settings, and unobtrusive HCI.

Future work will focus on refining signal processing algorithms for real-world air writing scenarios and addressing data inconsistencies arising from intuitive sensor placement. Additional longitudinal studies are needed to validate both writer identification and inter-subject character classification over time. On the hardware front, miniaturizing the sensing unit, particularly for fully wireless integration [46], as well as developing in-house directional antenna designs to improve spatial selectivity and robustness to gross arm motion, could significantly enhance user comfort and enable smoother integration into existing wearable platforms.

Beyond performance considerations, muscle motion data enables writer identification. Therefore, de-identification techniques can be useful to protect privacy when users do not wish to reveal their identity in certain scenarios. Our results confirm that biometric signals, such as muscle motion, encode identity traits similar to handwriting. By masking the specific waveform features responsible for personal identification, future systems can possibly preserve air writing functionality while safeguarding personal identity.