A Review: Radar Remote-Based Gait Identification Methods and Techniques

Abstract

Human identification using gait as a biometric feature has gained significant attention in recent years, showing notable advancements in medical fields and security. A review of recent developments in remote radar-based gait identification is presented in this article, focusing on the methods used, the classifiers employed, trends and gaps in the literature. Particularly, recent trends highlight the increasing use of Artificial Intelligence (AI) to enhance the extraction and classification of features, while key gaps remain in the area of multi-subject detection. In this paper, we provide a comprehensive review of the techniques used to implement such systems over the past 7 years, including a summary of the scientific publications reviewed. Several key factors are compared to determine the most suitable radar for remote gait-based identification, including accuracy, operating frequency, bandwidth, dataset, range, detection, feature extraction, size and number of features extracted, multiple subject detection, radar modules used, AI used and their properties, and the testing environment. Based on the study, it was determined that Frequency-Modulated Continuous-Wave (FMCW) radars were more accurate than Continuous-Wave (CW) radars and Ultra-Wideband (UWB) radars in this field. Despite the fact that FMCW is the most closely related radar to real-world scenarios, it still has some limitations in terms of multi-subject identification and open-set scenarios. In addition, the study indicates that simpler AI techniques, such as Convolutional Neural Network (CNN), are more effective at improving results.

1. Introduction

Throughout history, biometrics has been an essential tool for identifying individuals, beginning with simpler methods such as signatures and distinctive stamps, and evolving to more complex methods [1]. There is one objective that all of these techniques have, regardless of whether they were developed years ago or now, and that is to authenticate each individual and to prevent forgeries, thus ensuring security [2]. Nowadays, the techniques have become more complex and are divided into two categories: behavioral, which evaluate behaviors and manners, and physical, which evaluate a person’s physical characteristics [3]. Due to the need to directly measure part of a human body, such as fingerprints [4], palm prints [5], and iris prints [6], physical biometrics are more intrusive than psychological biometrics [7]. The use of behavioral biometrics can be less intrusive and independent of human interaction, as evidenced by the use of voice recognition [8] and gait analysis [9].

Biometrics based on gait, the way an individual walks, are an important aspect of security and medical fields due to the fact that they are not intrusive and can be monitored remotely [10]. However, this biometric faces some challenges, such as gait changes caused by injury, fatigue, or even aging that can affect the accuracy of the identification [11]. In order to overcome these challenges, various approaches have been developed, including the use of sensors such as Laser Imaging Detection and Ranging (LiDAR) sensors [12], normal cameras [13], thermal cameras [14], acoustic sensors [15], or the use of radars, such as Wireless Fidelity (Wi-Fi) radars [16], Long Range (LoRa) radars [17], Continuous Wave (CW) radars [18], Ultra-Wideband (UWB) radars [19] and Frequency Modulated Continuous Wave (FMCW) radars [20].

Among these methods, three radar techniques stand out, including CW radar, UWB radar, and FMCW radar, since they are capable of operating under adverse conditions such as low visibility [21] as well as respecting each individual’s privacy concerns by not taking photographs of the individual [22]. Due to their capability to capture gait information, even the smallest movements, through Doppler and micro-Doppler effects [23], these technologies are suitable for remote identification [24].

As any other technology, there are strengths and limitations. As a result of their simplicity, CW radars can only indirectly measure the target’s velocity [25,26]. It is already possible to measure both velocity and range from a target using UWB radar [27], and one of the best characteristics of UWB radars is their high resolution because of their high bandwidth [28]. However, their signal processing is more complex to implement due to the large amount of data to be processed [29]. The FMCW radar, which is a variation of the CW radar [30], modulates the signal frequency continuously to enable both range and velocity measurements [31]. Despite these advantages, the choice of the ideal technology for gait-based identification is still being debated and depends on several factors, including accuracy, achieved range, or the ability to identify multiple subjects.

An analysis of the literature comparing CW, UWB, and FMCW radars for gait-based identification is presented in this work. This comparison includes accuracy, operating frequency, bandwidth, dataset, range, detection, feature extraction, size and number of features extracted, multiple subject detection, radar modules used, the Artificial Intelligence (AI) used, and their properties of each radar type. The findings suggest that the FMCW radar is the most suitable choice for gait-based identification. In addition to the increased use of gait as a biometric indicator in various applications, this review is necessary due to the lack of consensus on which radar technology provides the most balance between accuracy, implementation complexity, and other factors. Additionally, there are no recent reviews that comprehensively evaluate the use of radar for gait-based identification. Furthermore, this study examines the State-of-the-Art (SOTA) of FMCW technology, including the AI techniques that achieve higher accuracy in addition to the factors discussed above.

In this paper, four sections are presented. Section 1 introduces biometric identification based on gait, highlighting radar technologies and describing what is studied. Section 2 presents a literature review for each of the CW, UWB, and FMCW radars. Section 3 discusses and concludes the most appropriate type of radar to use for gait-based identification. Finally, Section 4 summarizes the findings and suggests future research directions.

2. Gait-Based Identification

This section reviews the several studies on CW, UWB, and FMCW that have been conducted to acquire and process gait characteristics for biometric identification. Each subsection presents key information from the reviewed research, including the methodology and results. Moreover, a table is provided at the end of each subsection that compares studies using different radar types, namely

Table 1. Comparison of previous studies for biometric identification using CW radar.

Ref.	Date	Frequency [GHz]	Dataset Population	Range [m]	Detection	Features Extracted	Size and Features Dimensions	Multiple Subject Detection	Radar Module	AI Used	AI Properties	Accuracy	Environment
[32]	2017	-	25 M/4 F 18–47 YO	-	BI HM	Micro-Doppler signatures	Spectrograms length: 192 (dataset A) 39 time bins (dataset B) Time frame: 1.25 s	YES	-	LSTM-RNN	ADAM Optimizer Epochs: 100 Learning rate: $1\times {10}^{-3}$ Batch Size: 512	89.1% (dataset A: 192 steps) 87.07% (dataset B: 39 steps)	-
[33]	2018	24	12 M/12 F 24–28 YO 157–186 cm 48–75 kg	0–10	BI HM	Micro-Doppler signatures	Each person spectrograms: 4000 Size input spectrogram: $227\times 227$	NO	IVS-179	DCNN (AlexNet)	Caffe Network Learning rate: $1\times {10}^{-4}$ Batch Size: 32 (training) 16 (testing) Weight decay: $1\times {10}^{-4}$ Hidden nodes: 406	97.1% (4 S) 90.9% (6 S) 89.1% (8 S) 85.6% (10 S) 77.4% (12 S) 77.6% (16 S) 68.9% (20 S)	Outdoor (Walking area clean)
[34]	2019	25	17 M/5 F 162–195 cm 54–115 kg	3–10	BI HM	Micro-Doppler signatures	Size input image: $256\times 256\times 3$	NO	-	CNN (ResNet-50)	ADAM Optimizer	98%	Treadmill
[18]	2020	24	3 M/4 F 20–25 YO 160–175 cm 46–70 kg	3–10	BI HM	Micro-Doppler signatures	Features from LSTM: 2048 Features from CNN+RNN: 2048	NO	IVS-179	CNN+RNN LSTM	ADAM Optimizer Epochs: 100 Learning rate: $1\times {10}^{-4}$ Batch Size: 16	99% (validation set) 90% (test set)	Corridor (Walking area clean)
[35]	2021	10	16 M/6 F 21–55 YO 167–207 cm	3–25	BI HM	Micro-Doppler signatures	Spectrogram time duration: 1.25 s Total spectrograms: 12,803 Size input spectrogram: $128\times 192$	NO	-	DCNN (VGG-16)	ADAM Optimizer Epochs: 500 Learning rate: $1\times {10}^{-4}$ Learning rate decay: $1\times {10}^{-5}$ Mini Batches size: 32	93.5%	Outdoor (Walking area clean)
[36]	2022	5.8	6 M/4 F 23–33 YO 156–185 cm 56–86 kg	3–25	BI HM	Torso Doppler signals	Total samples: 1200 Each person samples: 120 Test samples size: $320\times 420$	NO	SDR-KIT 580B	CPCAN-3	SVM classifier Loss function: hinge loss	92.3%	Corridor (walking area clean)
[37]	2023	24	22 M/3 F Avg. 22.5 YO	4–12	BI HM	Gait time-velocity images	Window length: 32 samples (53.3 ms) Total images: 2625	NO	BSS-110	CNN (ResNet-16)	Learning rate: $1\times {10}^{-2}$–$3\times {10}^{-4}$ Batch Size: 64 Loss function: cross-entropy	99.1%	Outdoor: Walkway (Walking area clean)

M—Male, F—Female, S—Subjects, YO—Years Old; BI—Biometric Identification, HM—Human Motion; - Information not available.

for studies on CW radar, Table 2 for studies on UWB radar, and Table 3 for studies on FMCW radar.

2.1. Identification Using Continuous-Wave Radar

A milestone in the evolution of radar systems was the CW radar. With this type of radar, micro-Doppler signatures are captured, and Deep Learning (DL) techniques are employed to identify the person. Between the studies conducted for this type of radar, Klarenbeek et al. [32], combined a Recurrent Neural Network (RNN) model with Long- Short Term Memory (LSTM) for the classification of human gait radar signatures. Using CW radar on X-band, this method achieved an accuracy of 89.1%, outperforming traditional classifiers such as Convolutional Neural Network (CNN), Principal Component Analysis (PCA), and Support Vector Machine (SVM) methods. A key advantage of the RNN-LSTM model is its ability to handle different observation times, resulting in improved performance when using longer observation times.

Cao et al. [33] introduced a system based on Deep Convolutional Neural Network (DCNN). As a result of using Short-Time Fourier Transform (STFT) to extract features, the method reported an average accuracy of 97.1% in identifying four individuals and an average accuracy of 68.9% in identifying a group of twenty individuals, demonstrating a robust anti-noise capability. Limitations of the system include the variability in individual motion patterns and the reduction in effectiveness in Non-Line-of-Sight (NLOS) conditions.

Meanwhile, Abdulatif et al. [34] investigated the effect of human body characteristics on their measured data. As a result of the use of Combining Convolutional Autoencoders (CAE) and Residual Network (ResNet) with 50 layers, their method achieved 98% accuracy with a low Signal Noise Ratio (SNR), demonstrating their resistance to noise. According to [18], further refinement of gait recognition was achieved by integrating the LSTM, CNN and RNN architectures, which provided both long-term stability and recognition accuracy of 99% on the validation set and 90% on the test set, respectively.

The use of Visual Geometry Group (VGG), a DCNN with 16 layers, by Papanastasiou et al. [35] is another significant advancement, which, combined with a Uniform Manifold Approximation and Projection (UMAP), a dimensionality reduction method to validate the results, achieved an accuracy of 93.5%. The separation of micro-Doppler signals from limb movements from those generated by torso by Qiao et al. [36] increased the effectiveness on identification with a Three-layer Convolutional Principal Component Analysis Network (CPCAN-3). Lastly, in [37], it was reported that STFT methods are comparable to higher-resolution methods, achieving an accuracy of 99.1%, suggesting that simplicity and robustness often outperform complexity. Based on the parameters used, the data acquired, and the results obtained, Table 1 compares the research conducted with CW radar in the previous years.

2.2. Identification Using Ultra-Wideband Radar

As a result of UWB radar technology, significant advancements in biometric identification have been made, showing various approaches in signal processing, feature extraction, and classifiers to achieve better accuracy [38].

Table 2. Comparison of previous studies for biometric identification using UWB radar.

Ref.	Date	Frequency [GHz]	Dataset Population	Range [m]	Detection	Features Extracted	Size and Features Dimensions	Multiple Subject Detection	Radar Module	AI Used	AI Properties	Accuracy	Environment
[39]	2017	3.1–5.6	6 M/2 F 164–178 cm	-	BI HM BF	Raw Data	Scenarios tested: 384 Signature size: 61 frames	NO	NVA-R640	SVM	Also tested: RF LR KNN NN	88.15%	Over the entrance of a room
[40]	2019	3.2–5.4	8 M/7 F 154–179 cm	0–10	BI HM	Micro-Doppler signatures	-	NO	P410 RCM P410 MRM	No use	-	-	Anechoic chamber Laboratory
[41]	2019	3.993	8 M/4 F 19–44 YO 155–179 cm 56–95 kg	Wearable	BI HM	Interdistances between sensors	Walking time: 60 s Vector size elements: 72 Features functions: 12	NO	DecaWave EVB1000	subspace kNN weighted kNN bagged tree ESD SVM	-	96.9% (ESD) 96% (bagged tree) 95% (subspace kNN) 93% (SVM) 88% (weighted kNN)	Controlled
[42]	2019	4.3	4 S 23–25 YO 172–192 cm 65–90 kg	0–5	BI HM	Micro-Doppler spectograms	Samples per motion per person: 50 Spectrograms per motion per person: 50 (training set) Spectrograms per motion per person: 100 (testing set) Size spectrogram: $100\times 100$ Test time: under 1 ms	NO	Builded	MS-CNN	ADAM Optimizer Learning rate: $1\times {10}^{-3}$ Batch Size: 256 Max iteration: 1000	96.8% (walking) 84.8% (other activities)	Room (Walking area clean)
[19]	2019	4.3	10 M/5 F 160–187 cm 50–100 kg	5	BI HM	Micro-Doppler spectograms	Spectrograms per motion per person: 3000 Total spectrograms training set: 22,500 Total spectrograms testing set: 7500 Size spectrogram: $100\times 100$	NO	-	CNN	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Batch Size: 256	95.21% (running) 94.41% (walking) 82.93% (other activities)	Room (Walking area clean)
[43]	2020	3.1–5.3	6 S	3.5–7	BI HM	Limb Doppler signals	Measurement time: 8 s	NO	PulsON® 400	CNN	Stochastic gradient descent Learning rate: $1\times {10}^{-3}$	93.3%	Room (Walking area clean)
[44]	2021	3.2–5.4	13 M/11 F 153–179 Ccm	3	BI HM BF	Backscattered energies Knee angles	-	NO	P410 MRM	No use	-	-	Anechoic chamber Laboratory
[45]	2022	3.45–5.15	9 S	0–10	BI HM	Time-Doppler spectrograms	Total spectrograms: 20,494 Size spectrogram: $28\times 28$	NO	-	OR-CED	Stochastic gradient descent Learning rate: $1\times {10}^{-3}$ Epochs: 60 Mini Batches size: 32	96.17% (closed-set) 90.36% (open-set)	Room (Walking area clean)
[46]	2022	1.7 (bandwidth)	4 M/5 F	0–10	BI HM	Micro-Doppler signatures	Total spectrograms: +90,000 Echo segments: 720 Size spectrogram: $224\times 224$	NO	Builded	G-SAC	Stochastic gradient descent Learning rate: $1\times {10}^{-3}$ Epochs: 60 Batches size: 16	91.62%	Room (Walking area clean)
[47]	2023	3.45–5.15	10 M/5 F 160–187 cm 50–100 kg	-	BI HM	Micro-Doppler spectrograms	Spectrogram time duration: 1 s Total training spectrograms: 108,000 Validation and testing spectrograms: 36,000 Size spectrogram: $120\times 120$	NO	-	MS-CNN	ADAM Optimizer Epochs: 600 Learning rate: $1\times {10}^{-4}$ Batch size: 512	97% (walking) 85% (all tested actions)	-
[38]	2023	6–8.5	5 M/6 F 23–28 YO 160–187 cm 51–85 kg	2–5	BI HM	Raw Data	Size: $200\times 543$	NO	Xethru X4M03	MLRT (based on CNN)	Time-distributed CNN Output nodes: 13 (identification) 2 (fall-detection) Loss function: cross-entropy	98.7% (identification) 96.5% (fall-detection)	Room (Walking area clean)
[48]	2024	3.1–4.8	4 M/5 F 163–187 cm 52–79 kg	0–10	BI HM	Micro-Doppler signatures	Total spectrograms: +90,000 Echo segments: 720 Size spectogram: $224\times 224$	NO	Builded	DCNN (ResNet-50)	Batch Stochastic gradient Epochs: 30 Learning rate: $1\times {10}^{-2}$ Batch size: 32	61.46% (open-set)	Room (Walking area clean)

M—Male, F—Female, S—Subjects, YO—Years Old; BI—Biometric Identification, HM—Human Motion, BF—Body Figure; - Information not available.

Table 2. Comparison of previous studies for biometric identification using UWB radar.

Ref.	Date	Frequency [GHz]	Dataset Population	Range [m]	Detection	Features Extracted	Size and Features Dimensions	Multiple Subject Detection	Radar Module	AI Used	AI Properties	Accuracy	Environment
[39]	2017	3.1–5.6	6 M/2 F 164–178 cm	-	BI HM BF	Raw Data	Scenarios tested: 384 Signature size: 61 frames	NO	NVA-R640	SVM	Also tested: RF LR KNN NN	88.15%	Over the entrance of a room
[40]	2019	3.2–5.4	8 M/7 F 154–179 cm	0–10	BI HM	Micro-Doppler signatures	-	NO	P410 RCM P410 MRM	No use	-	-	Anechoic chamber Laboratory
[41]	2019	3.993	8 M/4 F 19–44 YO 155–179 cm 56–95 kg	Wearable	BI HM	Interdistances between sensors	Walking time: 60 s Vector size elements: 72 Features functions: 12	NO	DecaWave EVB1000	subspace kNN weighted kNN bagged tree ESD SVM	-	96.9% (ESD) 96% (bagged tree) 95% (subspace kNN) 93% (SVM) 88% (weighted kNN)	Controlled
[42]	2019	4.3	4 S 23–25 YO 172–192 cm 65–90 kg	0–5	BI HM	Micro-Doppler spectograms	Samples per motion per person: 50 Spectrograms per motion per person: 50 (training set) Spectrograms per motion per person: 100 (testing set) Size spectrogram: $100\times 100$ Test time: under 1 ms	NO	Builded	MS-CNN	ADAM Optimizer Learning rate: $1\times {10}^{-3}$ Batch Size: 256 Max iteration: 1000	96.8% (walking) 84.8% (other activities)	Room (Walking area clean)
[19]	2019	4.3	10 M/5 F 160–187 cm 50–100 kg	5	BI HM	Micro-Doppler spectograms	Spectrograms per motion per person: 3000 Total spectrograms training set: 22,500 Total spectrograms testing set: 7500 Size spectrogram: $100\times 100$	NO	-	CNN	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Batch Size: 256	95.21% (running) 94.41% (walking) 82.93% (other activities)	Room (Walking area clean)
[43]	2020	3.1–5.3	6 S	3.5–7	BI HM	Limb Doppler signals	Measurement time: 8 s	NO	PulsON® 400	CNN	Stochastic gradient descent Learning rate: $1\times {10}^{-3}$	93.3%	Room (Walking area clean)
[44]	2021	3.2–5.4	13 M/11 F 153–179 Ccm	3	BI HM BF	Backscattered energies Knee angles	-	NO	P410 MRM	No use	-	-	Anechoic chamber Laboratory
[45]	2022	3.45–5.15	9 S	0–10	BI HM	Time-Doppler spectrograms	Total spectrograms: 20,494 Size spectrogram: $28\times 28$	NO	-	OR-CED	Stochastic gradient descent Learning rate: $1\times {10}^{-3}$ Epochs: 60 Mini Batches size: 32	96.17% (closed-set) 90.36% (open-set)	Room (Walking area clean)
[46]	2022	1.7 (bandwidth)	4 M/5 F	0–10	BI HM	Micro-Doppler signatures	Total spectrograms: +90,000 Echo segments: 720 Size spectrogram: $224\times 224$	NO	Builded	G-SAC	Stochastic gradient descent Learning rate: $1\times {10}^{-3}$ Epochs: 60 Batches size: 16	91.62%	Room (Walking area clean)
[47]	2023	3.45–5.15	10 M/5 F 160–187 cm 50–100 kg	-	BI HM	Micro-Doppler spectrograms	Spectrogram time duration: 1 s Total training spectrograms: 108,000 Validation and testing spectrograms: 36,000 Size spectrogram: $120\times 120$	NO	-	MS-CNN	ADAM Optimizer Epochs: 600 Learning rate: $1\times {10}^{-4}$ Batch size: 512	97% (walking) 85% (all tested actions)	-
[38]	2023	6–8.5	5 M/6 F 23–28 YO 160–187 cm 51–85 kg	2–5	BI HM	Raw Data	Size: $200\times 543$	NO	Xethru X4M03	MLRT (based on CNN)	Time-distributed CNN Output nodes: 13 (identification) 2 (fall-detection) Loss function: cross-entropy	98.7% (identification) 96.5% (fall-detection)	Room (Walking area clean)
[48]	2024	3.1–4.8	4 M/5 F 163–187 cm 52–79 kg	0–10	BI HM	Micro-Doppler signatures	Total spectrograms: +90,000 Echo segments: 720 Size spectogram: $224\times 224$	NO	Builded	DCNN (ResNet-50)	Batch Stochastic gradient Epochs: 30 Learning rate: $1\times {10}^{-2}$ Batch size: 32	61.46% (open-set)	Room (Walking area clean)

M—Male, F—Female, S—Subjects, YO—Years Old; BI—Biometric Identification, HM—Human Motion, BF—Body Figure; - Information not available.

In 2017, Ghassem Mokhtari et al. [39] developed a system using UWB and Novelda sinuous antennas for identifying subjects based on raw data captured in the bandwidth of $3.1 \mathrm{GHz}$ to $5.6 \mathrm{GHz}$. In the study, some AI techniques were combined with PCA resulting in a classification accuracy to 88.15%. Algorithms such as SVM, Random Forest (RF), Logistic Regression (LR), K Nearest Neighbour (KNN), and Neural Network (NN) were used. This demonstrated their ability to extract signatures from basic movements with ease. These methods outperformed technologies that utilize Passive Infrared (PIR) sensors, wireless networks, and ground vibration sensors. Two years later, Soumya Prakash Rana et al. [40] developed an Impulse Radio Ultra-Wideband (IR-UWB) radar with monostatic antennas to capture micro-Doppler signatures. They applied the first three-dimensional approach based on spherical trigonometry and STFT to identify persons based on their gait. The study analyzed various elevation angles and azimuth angles, concluding that as the range is increased, the SNR decreases logarithmically.

Alessio Vecchio et al. [41] took a different approach by using wearable UWB sensors; this method involves measuring the distance between sensors to verify that accuracy increases when using multiple sensors in specific positions on the body. In the same year, in [42], Yue Lang et al. studied individual identification through the analysis of different movement patterns using horn antennas to capture data from four subjects. Movements such as boxing, crawling, creeping, jumping, running, and walking were tested. To avoid overadjustments in small datasets, a Multi-Scale Convolutional Neural Network (MS-CNN) was selected as the classifier. As a result, the accuracy rate was 96.8% in walking activities and 84.8% as the average across all activities. The identification of individuals based on different movement patterns was improved by Yang Yang et al. [19], who examined micro-Doppler signatures from fifteen subjects, with performance for running of 95.21%, for walking of 94.41%, and for other activities of 82.93%.

As of 2020, Takuya Sakamoto et al. [43] achieved 93.3% accuracy in analyzing the limb Doppler signals from six individuals using a UWB radar operating at $4.2 \mathrm{GHz}$ with a double ridge horn. This study demonstrated the importance of optimizing the input data size to reduce the learning time and avoid overfitting. Furthermore, Soumya Prakash Rana et al. [44] evaluated the knee angles in normal and abnormal gait from twenty-four subjects by comparing their efficiency with the Kinect system.

In 2022, Yang Yang et al. [45] proposed the Open-Set Classifier with Contrastive Embedding and Detection (OR-CED) for scenarios in which the number of training categories and the number of test categories differ, indicating an open-set test. If the training and test categories are the same, then a closed-set test is present. By using a classifier that utilizes a learning method to distinguish between hard positive and hard negative samples, they achieved 96.17% accuracy on closed-set scenarios and 90.36% accuracy on open-set scenarios. As part of another study conducted by the same researchers [46], the Guided Subspace Alignment under the Class-aware condition (G-SAC) model was implemented to measure the micro-Doppler signatures of nine subjects. Following the implementation of a Neighborhood Component Analysis (NCA), the researchers created an intrinsic feature subspace from which they could extract similarity between normal and disguised gaits. It was demonstrated that both supervised and unsupervised constraints could be used to ensure consistency in class-aware data distributions. This demonstrates the effectiveness of the method in identifying and validating multiple targets.

As another advancement, Yuan He et al. [47] implemented a MS-CNN model based on the contrastive learning framework. Using the multiscale feature extraction and image information interaction strategy, they were able to surpass previous studies, even when the SNR was low. Based on the results of the study, walking accuracy was 97% and the average of all activities tested was 85%, which is better than the studies performed in 2019.

Finally, advancements in the use of temporal and spatial characteristics were achieved through the Multi-task Learning Radar Transformer (MLRT) method by Xikang Jiang et al. [38]. This method integrated Kalman filters with mechanisms based on transformers, resulting in a 98.7% accuracy rate for identifying individuals. Additionally, the system was also developed to detect falls, achieving an accuracy rate of 96.5%. In addition, Yang Yang et al. [48] discussed how their Pseudoinvariant-features Separation for Domain Generalization (PSDG) model could be generalized to address the challenge of identifying disguised gaits in open scenarios. This approach resulted in an increase in accuracy of 5.51% when compared with other SOTA domain generalization techniques. In Table 2, the experimental results of each of the previous mentioned works are summarized.

2.3. Identification Using Frequency Modulated Continuous Wave Radar

While sharing similarities with CW radar, FMCW radar appears to overcome certain limitations such as the inability to measure range by integrating frequency modulation to enable distance estimation [31]. Xin Yang et al. [49] achieved 97.7% accuracy identifying a single user by using a CNN-based model that relies on lower limb motion and gait segmentation via silhouette. However, performance decreased to 91% when identifying two users and to 74% when identifying four users. A Deep Temporal Convolutional Neural Networks (DTCNN) was proposed in the same year by Addabbo et al. [50] to classify the micro-Doppler signatures with optimized temporal windows, resulting in the highest accuracy of 94.9% in all subjects. Also, Ni et al. [51] demonstrated 96.73% accuracy with ResNet-50 in a controlled environment.

Baori Zhou et al. [52] investigated the performance of CNNs on range-Doppler heat maps by testing algorithms such as AlexNet, VGGNet, GoogLeNet, and ResNet. With ResNet, they achieved the best accuracy, demonstrating a correlation between network depth and accuracy. Muhammad et al. [53] presented “gait cubes”, combining micro-Doppler and micro-Range signatures to create a CNN-based identification system capable of 96.1% accuracy, requiring minimal training data. In the following year, in 2021, Pegoraro et al. [54] used DCNNs with Inception Blocks (IBs) to investigate identification in complex environments such as corridors and real-life laboratory environments. They tested multi-user scenarios, obtaining 98.27% for four users at a time. As a validation result of Temporal Convolutional Network (TCN) in public datasets, Ref. [55] achieved an accuracy rate of 98.4% on a small dataset. However, the accuracy decreased as the dataset size increased.

The importance of novel loss functions and feature embeddings was highlighted in [56,57]. The Deep Discriminative Representation Network (DDRN) with Cosine Margin (CM) loss and the MS-CNN achieved accuracy rates of 95.94% and 88.57%, respectively. Xiang et al. [58] applied a Task Division and Multi-Task Learning (TD-MTL) module for pedestrian identification. Using a 25 min test set, they evaluated the model performance separately for pedestrian identification and validation. They achieved an accuracy rate of 87.63% for identifying five pedestrians and 80.78% for the validation set. The MGait developed in [59] overcame trajectory constraints. The study utilized a DCNN-based open-set classification approach for multi-user identification and intruder detection. MGait achieved 98.27% and 89.73% accuracy for two and five users, respectively.

Biometric fusion plays an important role in improving accuracy. A combination of vital signals and gait data was used by Alkasimi et al. [60], increasing accuracy over the use of a single biometric. Shi et al. [20] combined radar time-Doppler spectrograms with camera-derived Gait Energy Images (GEIs) to demonstrate the advantages of multimodal approaches for achieving robust identification. By integrating radar point clouds with micro-Doppler signatures, Ma and Liu [61] were able to improve system performance.

Table 3. Comparison of previous studies for biometric identification using FMCW radar.

Ref.	Date	Frequency [GHz]	Dataset Population	Range [m]	Detection	Features Extracted	Size and Features Dimensions	Multiple Subject Detection	Radar Module	AI Used	AI Properties	Accuracy	Environment
[49]	2020	77–81	10 S 21–34 YO 164–182 cm 45–74 kg	0–8	BI HM BF	Lower limb motion map	Size segments: $100\times 100$	YES	Texas Instruments AWR1642BOOST	CNN	Activation function: ReLU	97.7% (single: 10) 91% (multiple: 2) 84% (multiple: 3) 74% (multiple: 4)	(single: 5) Training != Test 78% (Lobby/Corridor) 74% (Corridor/Lobby) Training == Test 97.7% (Lobby) 95% (Corridor)
[50]	2020	77	5 M 23–32 YO 178–185 cm 60–99 kg	-	BI HM	Micro- Doppler signatures	Total time: 150 min Total training frames: 95,650 Total testing frames: 22,535 Total validation frames: 22,535	NO	-	DTCNN	Stochastic gradient descent Activation function: Mish Learning rate: $1\times {10}^{-1}$ Batch Size: 64 Weight decay: $1\times {10}^{-6}$ Loss function: cross-entropy	Top 1: 94.9%	2 Rooms
[51]	2020	76–78	13 M/7 F 23–45 YO 155–182 cm 48–83 kg	3–10.5	BI HM	Micro- Doppler signatures	Total time: 15 min Total training frames: 36,000 Total testing frames: 12,000	NO	Texas Instruments IWR1443 EVM	DCNN (ResNet-50)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$–$2\times {10}^{-3}$ Batch Size: 200	96.73%	Classroom (Walking area clean)
[52]	2021	77–81	5 S	-	BI HM	Range- Doppler heat map	Total frames: 3000 (1500 in each scenario) Total training heat maps: 12,000 Total testing heat maps: 3000	NO	Texas Instruments AWR1642BOOST	CNN	ADAM Optimizer Learning rate: $8\times {10}^{-5}$ (AlexNet) $1\times {10}^{-4}$ (VGGNet) $3\times {10}^{-4}$ (GoogLeNet) $1\times {10}^{-4}$ (ResNet) Epochs: 300 Batch Size: 62 Loss function: cross-entropy	96.9% (AlexNet) 97.6% (VGGNet) 97.7% (GoogLeNet) 97.9% (ResNet)	Indoor and Outdoor (Walking area clean)
[53]	2021	75.24–78.76	6 M/4 F 23–59 YO 160–174 cm 50–77 kg	2–10.91	BI HM	Micro- Doppler signatures Micro-Range signatures	Total gait cubes: 52,000	NO	Texas Instruments IWR1443BOOST	CNN	Stochastic gradient descent Learning rate: $1\times {10}^{-2}$ Batch Size: 80 Loss function: cross-entropy	96.1%	6 Areas (3 Open spaces and 3 Corridors)
[54]	2021	75–77	5 S	0–18	BI HM	Micro- Doppler signatures	Time bins spectrograms: 200	YES	INRAS	DCNN with IBs	Stochastic gradient descent Learning rate: $5\times {10}^{-3}$	(4 S) 97.96% (multiple: 2) 95.26% (multiple: 3) 98.27% (multiple: 4)	Room A: Corridor Room B: Laboratory (S: 2) Training != Test 96% (Room A/Room B)
[55]	2021	5.8	100 S 21–98 YO 149–198 cm	-	BI HM	Micro-Doppler spectrograms	-	NO	-	TCN	Loss function: cross-entropy Weight decay: $1\times {10}^{-6}$ 10 S: Nadam Optimizer Activation function: Swish Learning rate: $12\times {10}^{-2}$ Batch Size: 64	98.4% (10 S) 85.2% (50 S) 84.9% (100 S)	(Public dataset) Different environments
[56]	2021	76–78	20 S	3–10.5	BI HM	Micro-Doppler signatures	Total training samples: 18,000 Total testing samples: 6000 Size images: $224\times 224$	NO	Texas Instruments IWR1443 EVM	DDRN (ResNet-18)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Loss function: CM	95.94%	Controlled
[57]	2022	-	5 M 23–32 YO 178–185 cm 69–99 kg	-	BI HM	Range-Doppler map	Total time: 150 min Sample size input: $45\times 205$	NO	INRAS	MS-CNN	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Epochs: 300 Weight decay: $5\times {10}^{-4}$ Mini Batches size: 64	88.57%	-
[58]	2022	-	5 S	-	BI HM	Time-Doppler spectrograms	Size spectrogram: $45\times 205$	YES	-	MCL	Learning rate: $1\times {10}^{-3}$ Epochs: 500 Loss function: cross-entropy	87.63% (test set) 80.78% (validation set)	-
[59]	2022	75.85–79.15	5 M/5 F 23–33 YO 43–76 kg	3–23	BI HM	Micro-Doppler signatures	Images size: $224\times 224$	YES	-	DCNN (ResNet-18)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Loss function: large-margin Gaussian mixture	98.27% (multiple: 2) 95.60% (multiple: 3) 92.47% (multiple: 4) 89.73% (multiple: 5)	Corridor (Walking area clean)
[60]	2022	75–79	18 S 22–50 YO Avg. 171 cm Avg. 76 kg	-	BI HM VS	Doppler-Frame map Heart-Sound scalogram	Total time: 166 min Total samples: 28,800 Images size: $224\times 224$	NO	Texas Instruments AWR1642 EVM	DCNN (GoogLeNet)	Learning rate: $1\times {10}^{-4}$ Epochs: 30 Mini Batches size: 25 Gradient decay factor: 0.95 Squared gradient decay factor: 0.99	98% 58.7% (only Heart sounds) 96.2% (only Gait)	Corridor (Walking area clean)
[20]	2023	77	72 M/49 F 21–25 YO 155–187 cm 42–100 kg	-	BI HM BF	Micro-Doppler signatures GEIs	Total info: 80 pairs (GEIs and Time-Doppler spectrograms) GEIs size: $128\times 88$ Spectrogram size: $88\times 128$	NO	Texas Instruments AWR1843	DCNN	ADAM Optimization Learning rate: $1\times {10}^{-4}$	95.439% 87.198% (only GEIs) 54.232% (only radar) 92.178% (carrying a bag) 87.151% (wearing a coat)	Room (Walking area clean)
[62]	2023	77	5 M/4 F 18–35 YO 155–185 cm 45–110 kg	1–10	BI HM	Point Clouds	Total frames: +36,000	YES	Texas Instruments IWR1843BOOST	CNN + LSTM (PointNet++)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Batch Size: 16	96.75% (single: 9) 94.3% (multiple: 2)	95% (Laboratory) 80% (Corridor) 90% (Lobby)
[63]	2023	75.2–78.8	4 M/3 F 156–187 cm	0–8.24	BI HM	Micro-Doppler signatures Micro-Range signatures Range Maps	Total time: 50 min Total RDMs: 310,357 Frames per class per subject: 7400	NO	Texas Instruments AWR1443BOOST	LSTM	ADAM Optimizer Activation function: ReLU Learning rate: $1\times {10}^{-2}$ Epochs: 50 Batch Size: 512 Loss function: cross-entropy	93%	In-home (Real-life scenario)
[64]	2023	75.74–78.26	11 M/ 4F 21–41 YO 160–183 cm 51–75 kg	3–18	BI HM	4D Radar point cloud videos	Total frames: 45,000 Input size of the 4D RPCV: $50\times 64\times 4$	NO	Made by Texas Instruments	Spatial–Temporal Network (STN)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Epochs: 250 Batch Size: 32	94.44% (10 S) 90.76% (15 S)	Outdoor (Walking area clean)
[65]	2023	77	9 S	0.8	BI HM	Micro-Doppler signatures	-	NO	Texas Instruments IWR1443BOOST	CNN + TCN (ResNet)	GAM Optimization	(5 S) 96.44% (Identification) 98.29% (Gesture)	Lobby (Walking area clean)
[61]	2024	6–8.5	15 S 160–183 cm 51–75 kg	3–20	BI HM	Micro-Doppler signatures Radar point clouds	Size signatures: $128\times 126$ Size point cloud sequence: $50\times 4\times 48$	NO	Made by Texas Instruments	DCNN + TCN (PointNet)	ADAM Optimizer Learning rate: $5\times {10}^{-4}$ Epochs: 500 Loss function: Triplet loss with Center loss	81.65%	Room 3–16.5 m walking area clean 16.5–20 m walking area with desks
[66]	2024	77	4 M/4 F 25–33 YO 160–181 cm 45–85 kg	1.5–15.5	BI HM	Micro-Doppler spectrograms	Total frames: 40,000 (single) +5500 (multiple)	YES	Texas Instruments IWR1443 EVM	CNN (EfficientNet)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Batch Size: 32 Loss function: cross-entropy	98.5% (single: 8) 98.25% (multiple: 2) 97.30% (multiple: 3) 95.45% (multiple: 4)	Lobby (Walking area clean) and Corridor (Various reflections)
[67]	2024	77	2 S	-	BI HM	Micro-Doppler signatures Micro-Range signatures	Training samples: 560 Validation samples: 560 Testing samples: 234	YES	Texas Instruments AWR1642ISK-ODS	CNN+TCN	-	96.2% (multiple: 2)	(Scenario) Training != Test 67%

M—Male, F—Female, S—Subjects, YO—Years Old; BI—Biometric Identification, HM—Human Motion, BF—Body Figure, VS—Vital Signals; - Information not available.

Table 3. Comparison of previous studies for biometric identification using FMCW radar.

Ref.	Date	Frequency [GHz]	Dataset Population	Range [m]	Detection	Features Extracted	Size and Features Dimensions	Multiple Subject Detection	Radar Module	AI Used	AI Properties	Accuracy	Environment
[49]	2020	77–81	10 S 21–34 YO 164–182 cm 45–74 kg	0–8	BI HM BF	Lower limb motion map	Size segments: $100\times 100$	YES	Texas Instruments AWR1642BOOST	CNN	Activation function: ReLU	97.7% (single: 10) 91% (multiple: 2) 84% (multiple: 3) 74% (multiple: 4)	(single: 5) Training != Test 78% (Lobby/Corridor) 74% (Corridor/Lobby) Training == Test 97.7% (Lobby) 95% (Corridor)
[50]	2020	77	5 M 23–32 YO 178–185 cm 60–99 kg	-	BI HM	Micro- Doppler signatures	Total time: 150 min Total training frames: 95,650 Total testing frames: 22,535 Total validation frames: 22,535	NO	-	DTCNN	Stochastic gradient descent Activation function: Mish Learning rate: $1\times {10}^{-1}$ Batch Size: 64 Weight decay: $1\times {10}^{-6}$ Loss function: cross-entropy	Top 1: 94.9%	2 Rooms
[51]	2020	76–78	13 M/7 F 23–45 YO 155–182 cm 48–83 kg	3–10.5	BI HM	Micro- Doppler signatures	Total time: 15 min Total training frames: 36,000 Total testing frames: 12,000	NO	Texas Instruments IWR1443 EVM	DCNN (ResNet-50)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$–$2\times {10}^{-3}$ Batch Size: 200	96.73%	Classroom (Walking area clean)
[52]	2021	77–81	5 S	-	BI HM	Range- Doppler heat map	Total frames: 3000 (1500 in each scenario) Total training heat maps: 12,000 Total testing heat maps: 3000	NO	Texas Instruments AWR1642BOOST	CNN	ADAM Optimizer Learning rate: $8\times {10}^{-5}$ (AlexNet) $1\times {10}^{-4}$ (VGGNet) $3\times {10}^{-4}$ (GoogLeNet) $1\times {10}^{-4}$ (ResNet) Epochs: 300 Batch Size: 62 Loss function: cross-entropy	96.9% (AlexNet) 97.6% (VGGNet) 97.7% (GoogLeNet) 97.9% (ResNet)	Indoor and Outdoor (Walking area clean)
[53]	2021	75.24–78.76	6 M/4 F 23–59 YO 160–174 cm 50–77 kg	2–10.91	BI HM	Micro- Doppler signatures Micro-Range signatures	Total gait cubes: 52,000	NO	Texas Instruments IWR1443BOOST	CNN	Stochastic gradient descent Learning rate: $1\times {10}^{-2}$ Batch Size: 80 Loss function: cross-entropy	96.1%	6 Areas (3 Open spaces and 3 Corridors)
[54]	2021	75–77	5 S	0–18	BI HM	Micro- Doppler signatures	Time bins spectrograms: 200	YES	INRAS	DCNN with IBs	Stochastic gradient descent Learning rate: $5\times {10}^{-3}$	(4 S) 97.96% (multiple: 2) 95.26% (multiple: 3) 98.27% (multiple: 4)	Room A: Corridor Room B: Laboratory (S: 2) Training != Test 96% (Room A/Room B)
[55]	2021	5.8	100 S 21–98 YO 149–198 cm	-	BI HM	Micro-Doppler spectrograms	-	NO	-	TCN	Loss function: cross-entropy Weight decay: $1\times {10}^{-6}$ 10 S: Nadam Optimizer Activation function: Swish Learning rate: $12\times {10}^{-2}$ Batch Size: 64	98.4% (10 S) 85.2% (50 S) 84.9% (100 S)	(Public dataset) Different environments
[56]	2021	76–78	20 S	3–10.5	BI HM	Micro-Doppler signatures	Total training samples: 18,000 Total testing samples: 6000 Size images: $224\times 224$	NO	Texas Instruments IWR1443 EVM	DDRN (ResNet-18)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Loss function: CM	95.94%	Controlled
[57]	2022	-	5 M 23–32 YO 178–185 cm 69–99 kg	-	BI HM	Range-Doppler map	Total time: 150 min Sample size input: $45\times 205$	NO	INRAS	MS-CNN	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Epochs: 300 Weight decay: $5\times {10}^{-4}$ Mini Batches size: 64	88.57%	-
[58]	2022	-	5 S	-	BI HM	Time-Doppler spectrograms	Size spectrogram: $45\times 205$	YES	-	MCL	Learning rate: $1\times {10}^{-3}$ Epochs: 500 Loss function: cross-entropy	87.63% (test set) 80.78% (validation set)	-
[59]	2022	75.85–79.15	5 M/5 F 23–33 YO 43–76 kg	3–23	BI HM	Micro-Doppler signatures	Images size: $224\times 224$	YES	-	DCNN (ResNet-18)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Loss function: large-margin Gaussian mixture	98.27% (multiple: 2) 95.60% (multiple: 3) 92.47% (multiple: 4) 89.73% (multiple: 5)	Corridor (Walking area clean)
[60]	2022	75–79	18 S 22–50 YO Avg. 171 cm Avg. 76 kg	-	BI HM VS	Doppler-Frame map Heart-Sound scalogram	Total time: 166 min Total samples: 28,800 Images size: $224\times 224$	NO	Texas Instruments AWR1642 EVM	DCNN (GoogLeNet)	Learning rate: $1\times {10}^{-4}$ Epochs: 30 Mini Batches size: 25 Gradient decay factor: 0.95 Squared gradient decay factor: 0.99	98% 58.7% (only Heart sounds) 96.2% (only Gait)	Corridor (Walking area clean)
[20]	2023	77	72 M/49 F 21–25 YO 155–187 cm 42–100 kg	-	BI HM BF	Micro-Doppler signatures GEIs	Total info: 80 pairs (GEIs and Time-Doppler spectrograms) GEIs size: $128\times 88$ Spectrogram size: $88\times 128$	NO	Texas Instruments AWR1843	DCNN	ADAM Optimization Learning rate: $1\times {10}^{-4}$	95.439% 87.198% (only GEIs) 54.232% (only radar) 92.178% (carrying a bag) 87.151% (wearing a coat)	Room (Walking area clean)
[62]	2023	77	5 M/4 F 18–35 YO 155–185 cm 45–110 kg	1–10	BI HM	Point Clouds	Total frames: +36,000	YES	Texas Instruments IWR1843BOOST	CNN + LSTM (PointNet++)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Batch Size: 16	96.75% (single: 9) 94.3% (multiple: 2)	95% (Laboratory) 80% (Corridor) 90% (Lobby)
[63]	2023	75.2–78.8	4 M/3 F 156–187 cm	0–8.24	BI HM	Micro-Doppler signatures Micro-Range signatures Range Maps	Total time: 50 min Total RDMs: 310,357 Frames per class per subject: 7400	NO	Texas Instruments AWR1443BOOST	LSTM	ADAM Optimizer Activation function: ReLU Learning rate: $1\times {10}^{-2}$ Epochs: 50 Batch Size: 512 Loss function: cross-entropy	93%	In-home (Real-life scenario)
[64]	2023	75.74–78.26	11 M/ 4F 21–41 YO 160–183 cm 51–75 kg	3–18	BI HM	4D Radar point cloud videos	Total frames: 45,000 Input size of the 4D RPCV: $50\times 64\times 4$	NO	Made by Texas Instruments	Spatial–Temporal Network (STN)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Epochs: 250 Batch Size: 32	94.44% (10 S) 90.76% (15 S)	Outdoor (Walking area clean)
[65]	2023	77	9 S	0.8	BI HM	Micro-Doppler signatures	-	NO	Texas Instruments IWR1443BOOST	CNN + TCN (ResNet)	GAM Optimization	(5 S) 96.44% (Identification) 98.29% (Gesture)	Lobby (Walking area clean)
[61]	2024	6–8.5	15 S 160–183 cm 51–75 kg	3–20	BI HM	Micro-Doppler signatures Radar point clouds	Size signatures: $128\times 126$ Size point cloud sequence: $50\times 4\times 48$	NO	Made by Texas Instruments	DCNN + TCN (PointNet)	ADAM Optimizer Learning rate: $5\times {10}^{-4}$ Epochs: 500 Loss function: Triplet loss with Center loss	81.65%	Room 3–16.5 m walking area clean 16.5–20 m walking area with desks
[66]	2024	77	4 M/4 F 25–33 YO 160–181 cm 45–85 kg	1.5–15.5	BI HM	Micro-Doppler spectrograms	Total frames: 40,000 (single) +5500 (multiple)	YES	Texas Instruments IWR1443 EVM	CNN (EfficientNet)	ADAM Optimizer Learning rate: $1\times {10}^{-4}$ Batch Size: 32 Loss function: cross-entropy	98.5% (single: 8) 98.25% (multiple: 2) 97.30% (multiple: 3) 95.45% (multiple: 4)	Lobby (Walking area clean) and Corridor (Various reflections)
[67]	2024	77	2 S	-	BI HM	Micro-Doppler signatures Micro-Range signatures	Training samples: 560 Validation samples: 560 Testing samples: 234	YES	Texas Instruments AWR1642ISK-ODS	CNN+TCN	-	96.2% (multiple: 2)	(Scenario) Training != Test 67%

M—Male, F—Female, S—Subjects, YO—Years Old; BI—Biometric Identification, HM—Human Motion, BF—Body Figure, VS—Vital Signals; - Information not available.

To ensure resilience against environmental changes, Dang et al. [62] integrated PointNet++ and Gated Recurrent Unit (GRU) into PGGait. The system extracts point clouds and uses them as inputs for the NN, achieving over 96.75% accuracy; it also showed that the accuracy of the system varied minimally with different walking directions. By using a cloud-based system with Multiple-Input Multiple-Output (MIMO) FMCW radar, Abedi et al. [63] demonstrated the possibility of performing real-time activity recognition by applying an LSTM-based model.

Micro-Doppler approaches, using 4D radar point clouds and Density-Based Spatial Clustering of Applications (DBSCAN) clustering, have been developed for tracing trajectories in outdoor environments. To eliminate noise and improve targeting, algorithms such as two-dimensional constant false alarm rate (2D-CFAR) and DBSCAN were employed in [64], achieving an accuracy of 94.4%. A network based on the Identity Recognition Network (IDRVT) method was developed by Ding et al. [65], which incorporated local, global, and temporal information using the combination of CNNs and TCNs. By using Gradient Norm Aware Minimization (GAM), the accuracy improved to 96.44%.

Finally, in 2024, Li et al. [66] developed MCGait, which can identify multiple users without limitations on walking trajectories using micro-Doppler signatures and an EfficientNet architecture. It achieved 98.5% accuracy in single-person identification and 98.25% accuracy in multisubject identification, even in scenarios with many reflections. Additionally, Huang et al. [67] demonstrated the generalization of micro-Doppler signature in a test scenario differing from the training scenario. When the test scenario matched the training data, the model achieved 96.2% accuracy. However, in a previously unseen scenario, performance decreased to 67% accuracy, highlighting the challenge of domain adaptation. To facilitate understanding of the discussed studies, Table 3 provides a clearer comparative analysis of all reviewed works.

3. Discussion

A review of the literature relating to CW, UWB, and FMCW radars for gait-based identification is presented in this section. The main objective is to summarize the key findings, compare these three technologies, and suggest which is the most suitable for this application. To summarize and compare the reviewed data, Table 1, Table 2 and Table 3 were utilized. These tables provided the data used to create the figures presented in this section.

Although all technologies are capable of capturing unique gait patterns, their structures play a critical role in their performance, which is the main factor in this field study. In general, the more accurate a system is, better suited it is for real-world applications. However, it should also consider other factors besides its accuracy.

Figure 1 shows the number of articles that have been published in this field. It is evident that FMCW radar has been studied for this application, exceeding the 50% mark in the number of research studies conducted. Less than half of the articles focused on UWB radar studies and CW radar studies, in which UWB radar was more commonly used than CW radar.

The operating frequency and bandwidths of the systems vary from one another, directly influencing the capability to detect targets and the range resolution [68]. As shown in Figure 2, CW radars typically operate within the $23 \mathrm{GHz}$ to $26 \mathrm{GHz}$ range, UWB radars within $3 \mathrm{GHz}$ to $6 \mathrm{GHz}$ range, and FMCW radars with a central frequency around $77 \mathrm{GHz}$. These values are due to the fact that researchers tend to use commercially available radar hardware as opposed to developing custom systems. Researchers who have studied UWB radar are the only ones who have constructed their own radars, as can be seen in [42,46,48]. Despite being more time-consuming, the development of custom radar systems is usually better as it allows researchers to design according to their needs and can potentially lead to cheaper systems. When purchasing a radar kit, a potential limitation on its resolution can be attributed to their bandwidth, which ranges between $1.7 \mathrm{GHz}$ and $2.5 \mathrm{GHz}$ for UWB radar kits and between $2 \mathrm{GHz}$ and $4 \mathrm{GHz}$ for FMCW radar kits.

In Figure 3, it can be seen that accuracy tends to increase with the increase in bandwidth. An increase in bandwidth results in an increase in resolution [69], which leads to an increase in accuracy. A number of other factors influence accuracy, such as the dataset diversity, the detection of multiple subjects, and the use of Machine Learning (ML) methods. Additionally, the graph indicates that studies with more participants are more likely to have higher accuracy. This is counterintuitive, as an increase in the number of participants usually introduces more variability, making accurate classification more difficult.

As shown in Figure 4, where all accuracies are taken into account from all studies, the accuracy tends to remain constant with increasing bandwidth. This situation can be considered normal, since it encompasses studies involving multiple subject detections as well as open-set scenarios, and studies examining the accuracy of activities other than walking. Based on the accuracy values in both figures, FMCW radar operates effectively at bandwidths above $2 \mathrm{GHz}$, and the UWB radar is often designed for use below a bandwidth of $2.5 \mathrm{GHz}$.

There has also been a trend observed in the radar operating range used by the researchers. Despite the fact that all radar types tend to not pass the $10 \mathrm{m}$ detection range, as shown in Figure 5, studies with CW radar achieved detection ranges of $25 \mathrm{m}$ in [35,36] and with FMCW radar detection ranges were $20 \mathrm{m}$ and $23 \mathrm{m}$ in [59,61], respectively. UWB radar has the lowest range, showing a maximum of $10 \mathrm{m}$, indicating that gait-based identification at longer distances remains a challenge. Their wide bandwidth may improve range resolution but disperses transmit power, resulting in reduced power spectral density. Consequently, signal attenuation increases, SNR decreases, and detection range is reduced [70]. As a result, there is a gap in the field concerning longer ranges, and it is a subject worth studying since only four works were able to extend beyond the $20 \mathrm{m}$ range.

As the range increases, the accuracy is expected to decrease due to signal attenuation or interference from the environment [71]. According to Figure 6, the majority of studies with lower ranges tend to have higher accuracy due to lower signal attenuation [72].

It is also important to consider both the size and variability of the datasets used to train the classifier as well as the number of subjects in the test group. When a classifier is trained on a large and diverse set of data, its robustness and generalization are generally enhanced, which allows it to better simulate real-world conditions. In contrast, smaller datasets may limit the classifier’s effectiveness in real-world scenarios, as is the case of [54], which achieved an accuracy of 98.27% with a training set of five subjects and a testing group of four subjects. A good example of study with a large dataset diversity is [60], which was able to achieve an accuracy rate of 98% with a dataset of 18 subjects. By merging gait information with heart sounds, extracted from a heart sound scanner, they were able to achieve this high level of accuracy.

This leads to the next important factor, the combination of biometrics. According to Figure 7, only a few studies have attempted to combine biometrics and radar, more specifically the FMCW radar studies. Ref. [60] combines Doppler-frame mapping with heart sounds scalograms to improve the accuracy in relation to the studies that use a single biometric. There is another investigation of combining biometrics in [20], where micro-Doppler signatures are combined with GEIs, as well as in [61], where micro-Doppler signatures are combined with 4D radar point cloud sequences.

Studies [20,60,61] have the same objective, which is to demonstrate that a combination of extracted features is more effective than using only one input in a classifier. However, Ref. [61] does not show this, obtaining a very low accuracy of 81.65%, due to challenges with cross-view conditions, resulting in a lower median value in Figure 8. According to Figure 8, we can conclude that the combination of biometrics is beneficial, improving the systems accuracy in most cases, but there may be scenarios in which this approach is not effective.

By analyzing the features extracted from the tables, it can be concluded that micro-Doppler signatures and spectrograms are the most frequently used features. However, it is important to note that these signatures may also include environmental noise and variations in the body. In addition, “gait cubes” are introduced in [53], which combine micro-Doppler, micro-Range, and temporal data to reduce the need of large training datasets.

These micro-Doppler signatures and spectrograms are the most relevant feature to be extracted to use in the classifier due to their ability to capture fine-grained motion patterns of different body patterns, such as the head, torso, or limbs. These features are also capable of capturing the unique characteristics of each individual, such as walking speed, stride length, and step frequency [66].

The features extracted represent the data to be fed into the classification model, and their size and number affect the system accuracy directly. It is common for features like micro-Doppler spectrograms and micro-Doppler signatures to have larger dimensions, such as $224\times 224$ or $227\times 227$, which provides a detailed description of human movements and at the same time increases computational demands and the risk of overfitting. Overall, the choice of the number and size of features depends on the environment and the application. When operating in controlled environments, higher dimensions can be supported to maximize accuracy, as in [56], while real-time systems can benefit from lower dimensions [63]. Therefore, it is necessary to establish a balance between the dimensions of the features and the computational efficiency of the model.

To establish an appropriate balance between feature dimensions and computational efficiency, the intended application must be considered. In real-time scenarios, models based on lower-dimensional features are generally preferred, as they enable faster processing. Conversely, architectures such as ResNet-50 or other DCNNs using input sizes of $224\times 224$ or higher tend to achieve superior accuracy, although the cost of computational demands increased.

In order to classify these features, machine learning methods are used. This illustrates the importance of incorporating AI into the system. Figure 9 illustrates the frequency that each AI technique is applied for gait-based identification. It can be concluded from Figure 9 that CNNs, DCNNs, and their variants are the most popular options in this field. The hybrid approaches of [18,65] combine the strengths of CNNs with temporal dependencies and spectrum analysis to improve performance. In this analysis, we observe that simpler methods, such as CNNs, are the most commonly used, although there is a growing increase in the use of DL, as in the case of hybrid methods and DCNNs.

Regarding AI properties, it can be seen that optimizers such as Adaptive Moment Estimation (ADAM) Optimizer, loss functions such as cross-entropy, and activation functions such as Rectified Linear Unit (ReLU) are frequently used. Hyperparameters, such as learning rates, batch sizes, and epochs, were adjusted according to the data complexity and feature dimensions to maximize performance.

It is also important to take into account the identification of multiple subjects when selecting an AI algorithm. According to Figure 10, it is still a topic that needs to be discussed. As of the time of writing this work, only eight works from the thirty-nine have been able to identify multiple subjects, none of which were from UWB radars.

According to Figure 11, when multi-subject detection is achieved, it commonly uses CNNs and DCNNs.

Researchers generally achieved high accuracy values using various methods for multi-user identification. Among these methods, segmenting individual steps and analyzing lower limb movements [49] or using Kalman filters in conjunction with DCNNs to achieve person tracking and identification [54] proved to be effective options. However, the number of simultaneously identified individuals did not exceed five subjects on studies with FMCW radar and two subjects with CW radar. Additionally, from Figure 11, it is evident that the accuracy decreases with an increase in the number of users due to an increase in signal interference and overlapping patterns, which make it more difficult for the algorithms to distinguish between individuals. Despite the achievement of multiple subject detection, the majority of the studies were conducted in controlled environments, challenging their applicability to real-world scenarios.

A possible solution to this may be the use of DCNN, as it is the AI technique with the highest accuracy and the only one capable of identifying five subjects simultaneously. Furthermore, the performance of DCNN-based models tends to remain consistent between single-subject and multi-subject scenarios, providing that the FMCW radar is capable of effectively separating individual subjects in space [73].

In this way, the environment has an impact on the measured accuracy. For example, if the test is conducted in a controlled environment or not, in the same location as training tests or not, and if there is a greater or lesser degree of signal reflection.

When analyzing the three tables, the studies were conducted in controlled environments, usually in empty rooms, lobbies, and corridors, lacking real-life testing. However, some studies attempted to determine whether the classification would perform as well in a testing environment different from the one used for training. This work was only made with FMCW radar, and according to the results shown in Figure 12, there is a decrease of 20.32% in accuracy when the training environment is different from the testing environment.

The presence of static and dynamic objects creates unwanted reflections and noise, which can distort the radar signal and affect the accuracy of the classifier. Studies usually take the data from environments that have no obstacles, such as corridors, lobbies, and empty rooms, to avoid these issues. However, in some studies, such as [62], the obstacles are present in the corners of the room, or in [66] the corridor tested has large windows and glass walls. This can generate unwanted reflections and ghost targets, simulating real-world scenarios.

In order to eliminate noise and unwanted reflections from the environment, the raw data signal undergoes a preprocessing stage before being applied to a classifier. For classification, this step is crucial to ensuring that only relevant features are extracted. For noise filtering, steps such as 2D-Fast Fourier Transform (FFT) and interference suppression are essential. There is usually a significant reduction in accuracy when the pre-processing module is omitted [64].

There is also the possibility of open-set scenarios, where the radar captures data from an intruder, a person who has not taken part in the training data for the classifier. According to [59], a DCNN-based classifier along with DBSCAN algorithm is used to identify the intruder, achieving 91.36% with one intruder per five persons, and decreasing as the number of intruders increases. An open-set scenario with UWB radar was also explored in [48], in which unknown gaits were evaluated, with DCNN, more specifically with ResNet-50, achieving 61.46% accuracy. It is important to consider that once systems are used in real-world scenarios, they cannot rely solely on trained datasets and must be capable of rejecting subjects unknown to the system.

Taking a closer look at Figure 13 and Figure 14, it is possible to verify that CW has the lowest median value of accuracy, indicating that this type of radar is the least effective for achieving gait-based identification. Possibly, this is the result of the simple radar architecture, which makes it impossible to measure the range (distance) and can only measure, indirectly, velocity. Thus, the CW radar is the worst possible option for gait-based identification.

Both FMCW and UWB radars have high median accuracy values, but FMCW has the highest. This can be attributed to the fact that, since UWB systems are more complex and expensive to implement when compared to FMCW, they see less use and are less researched, thus having a lower median. As discussed previously in this section, FMCW radars are capable of detecting multiple subjects and have longer ranges than UWB radars.

Considering the FMCW radar has higher accuracy, ability to measure both range and velocity, lower cost, and ease of implementation, it appears to be a promising option for gait-based identification.

In light of this, focusing on FMCW as the preferred type for gait-based identification, Figure 15 and Figure 16 illustrate how much a given AI technique is used for gait-based identification. Based on the analysis of both figures, CNNs, DCNNs, and their variants are the most popular choices in this area, as seen in Figure 9. This is possibly due to their ability to extract detailed spatial features more effectively than other methods [74]. The use of other methods, such DDRN and MCL, is unusually combined with this radar technology. There have also been very few hybrid approaches, having only combined DCNN with TCN [61], CNN with LSTM [62], and CNN with TCN [65,67]. Combining CNN or DCNN with TCN can be advantageous, since it integrates the spatial feature extraction capabilities of NN with the temporal modeling capabilities of TCN.

In addition, the accuracy of these two architectures, CNN and DCNN, is greater than that of all other methods, as shown in Figure 17 and Figure 18. Furthermore, the variability of the results in these two ML methods is lower than in others, indicating their consistency in achieving high accuracy values. As a result of this stability, CNNs and DCNNs appear to provide reliable performance on a wide variety of datasets and under diverse conditions. It is important to note that DCNNs are capable of very high performance, as illustrated in Figure 17, but they are also more complex, more time-consuming, and require a larger amount of data [75]. The median value of the DCNN decreased when considering all accuracy values, as illustrated in Figure 18. Considering this, DCNNs may be capable of achieving peak performance, but their accuracy may vary depending on specific training conditions or datasets. Although other techniques are also used in the area, the resources are concentrated on the two mentioned previously.

Overall, the literature review identifies some trends and gaps in the field of gait-based identification using radar. The trends are listed as follows:

• The FMCW radar technology is the subject of most research;
• An increase in bandwidth is associated with an increase in resolution, which increases accuracy;
• In AI, CNNs and DCNNs are the most commonly used techniques;
• Micro-Doppler signatures and spectrograms are usually inputs to the classifier;
• Using multimodal systems increases the accuracy of identification;
• The majority of the studies are conducted in controlled environments.

The gaps are as follows:

• Datasets with low diversity;
• Systems with a short range: few works have exceeded 20 m;
• Multimodal systems are rarely utilized;
• The number of works detecting multiple subjects is very low;
• In most cases, controlled environments are used for testing.

4. Conclusions

In this article, different radar technologies are compared, CW, UWB, and FMCW radar, to determine the best choice of radar technology for gait-based identification in real-world scenarios, focusing on performance and implementation complexity. An array of factors were considered in the analysis, including the accuracy of the identification, the operating frequency, the bandwidth, the dataset, the range, the detection, the feature extraction, the detection of multiple subjects, and the use of AI. Based on the results, it was found that the CW radar, once considered the most simple, presents the lowest accuracy value compared with the others radar systems. When comparing UWB and FMCW radars, FMCW is more accurate, simpler to implement, and can detect multiple subjects and achieve higher ranges. The use of AI techniques for biometric identification is another important aspect. Among the methods studied, CNNs and DCNNs are the most efficient, mainly in micro-Doppler signature features. In addition, the fusion of biometrics can enhance accuracy; however, this possibility has only been studied by a few researchers. Despite the improvements shown, most studies were conducted under controlled environments and with small datasets. Additionally, most of the research was conducted in closed-set scenarios. Further work can be undertaken to validate the systems on real-world scenarios, to test multi-subject identification with more than five users, to explore biometric fusions, and to develop algorithms for open-set scenarios. Emerging AI techniques such as self-supervised learning offer promising solutions to reduce the reliance on large datasets, which remains a limitation in current gait identification studies. Future research should investigate the integration of these methods to improve robustness, especially in open-set or multi-subject scenarios and under real-world variability. In summary, FMCW radar still has much potential to be explored and developed, but it has already demonstrated its advantages for applications such as gait-based identification.