Open Access
Add to BibSonomy
Abstract
Current perception and monitoring systems, such as human recognition, are affected by several environmental factors, such as limited light intensity, weather changes, occlusion of targets, and public privacy. Human recognition using radar signals is a promising direction to overcome these defects; however, the low signal-to-noise ratio of radar signals still makes this task challenging. Therefore, it is necessary to use suitable tools that can efficiently deal with radar signals to identify targets. Reservoir computing (RC) is an efficient machine learning scheme that is easy to train and demonstrates excellent performance in processing complex time-series signals. The RC hardware implementation structure based on nonlinear nodes and delay feedback loops endows it with the potential for real-time fast signal processing. In this paper, we numerically study the performance of the optoelectronic RC composed of optical and electrical components in the task of human recognition with noisy micro-Doppler radar signals. A single-loop optoelectronic RC is employed to verify the application of RC in this field, and a parallel dual-loop optoelectronic RC scheme with a dual-polarization Mach–Zehnder modulator (DPol-MZM) is also used for performance comparison. The result is verified to be comparable with other machine learning tools, which demonstrates the ability of the optoelectronic RC in capturing gait information and dealing with noisy radar signals; it also indicates that optoelectronic RC is a powerful tool in the field of human target recognition based on micro-Doppler radar signals.
Published by Optica Publishing Group under the terms of the Creative Commons Attribution 4.0 License. Further distribution of this work must maintain attribution to the author(s) and the published article's title, journal citation, and DOI.
1. INTRODUCTION
With the rapid evolution of novel Internet and mobile services, the need for intelligent perception scenarios such as automatic recognition of human targets is also developing rapidly. Intelligent perception and monitoring systems play an increasingly important role in the digital construction of modern cities, but these systems currently mainly use tools such as visual cameras based on visible light to collect target information. This method will be affected by many environmental factors, such as limited light intensity, changes in weather, and occlusion of targets; it is also inconvenient to set in privacy-sensitive areas and easily damaged by malicious intent. Therefore, improving the automatic perception- and intelligent-sensing capabilities of an environment is an important direction for the development of smart cities in the future.
Using radar detection can solve the aforementioned problems well, and there are many application scenarios, such as autonomous driving [1,2], border control [3], earthquake search [4] and rescue, and military search [5]. Target recognition based on radar signals utilizes the micro-Doppler effect. The micro-motion of the target, such as vibration and rotation, will cause additional frequency modulation on the Doppler frequency shift in the radar echo, that is, the micro-Doppler effect [6]. Different heights, body shapes, pace, and posture of human targets have different feedback on the Doppler frequency of the radar echo. Therefore, radar signals can be used to distinguish different human targets.
Some research work on the human recognition with the micro-Doppler (MD) radar signals have been reported [7–11]. In [7], seven men and six women walked on a treadmill located in front of the radar equipment; the authors proposed the use of K-mean clustering and ${ K}$-nearest neighbor methods to process the micro-Doppler features of the radar signals. The recognition accuracy rate of all subjects reached 87%, and the gender judgment accuracy rate was 92.4%. In [10], the authors constructed a system that achieved an accuracy of 88.6% in identifying five subjects, and the five subjects had a large difference in body size. In the aforementioned research work, the identified scenes are relatively simple, such as corridors or treadmills, and they are all radial movements, which are not close to reality.
In more complex testing scenarios, subjects are allowed to move around freely, and these testing scenarios are more meaningful and challenging [12–15]. In [12], the authors provided a data set of radar signals of five subjects randomly walking around a room and process the micro-Doppler signatures using a deep convolutional neural network (DCNN). The target recognition accuracy rate reached 73% when the training set and test set data were collected in different rooms. In [13], based on the data set in [12], the author proposed a DCNN method based on the initial block, which further improves the accuracy to 90%. However, the complexities of the employed methods are quite high; further, these methods are difficult to implement in hardware, and the training and testing time is long.
This paper aims to explore machine learning methods that can efficiently process radar micro-Doppler signatures. Reservoir computing (RC) is excellent at dealing with timing-related issues [16,17]. RC originates from the improvement of the recurrent neural network (RNN), which reduces the training complexity by randomly generating the connection matrix of the internal network, and it only needs to train the output connection weight matrix linearly. Meanwhile, the structure of random interconnection of nodes in RC can keep the input signal in the network for a long time, which overcomes the problem of memory fading in RNN [18]; thus, it can solve the problem of long time-series signals and high noisy signals. The hardware implementation methods of RC have been gaining increasing attention [19,20] because of their ability to perform high-speed, efficient, and real-time processing [21]. The hardware implementation of RC can be realized with pure electrical methods [22], optoelectronic methods [23–25], and all-optical methods [26,27]. The main difference among these methods lies in the categories of the components that make up the RC. The pure electrical method is designed with all-electrical devices in which the bandwidth is limited and the power consumption is high. Comparatively, the optoelectronic and all-optical RC schemes are favored thanks to their high bandwidth, fast processing speed, and low power consumption. However, the all-optical RC schemes have complex hardware requirements, and the stability is still limited. Therefore, the optoelectronic RC is considered a potential method to deal with the low SNR radar signal of human recognition.
In this paper, we innovatively explore the performance of the optoelectronic RC in the task of human target recognition based on radar signals utilizing numerical study, using the micro-Doppler feature as the input signal of a single-loop optoelectronic RC and a parallel dual-loop RC, which has been verified to have stronger performance in dealing with nonlinear signals [24]. We demonstrate the performance of the optoelectronic RC in processing noisy radar signals, indicating the potential of the optoelectronic RC in human target recognition tasks.
2. OPTOELECTRONIC RESERVOIR COMPUTING AND RADAR SIGNAL PREPROCESSING
A. Delay Feedback-Loop Based Reservoir Computing
The reservoir computing contains an input layer, hidden layers, and an output layer. The commonly used hardware implementation scheme is based on the structure of a nonlinear node in addition to a delay feedback loop. Figure 1 is a schematic diagram of the structure of delay feedback RC [22,23].
Fig. 1. Schematic diagram of the structure of delay feedback loop RC. ${{\boldsymbol W}_{{\textbf{in}}}}$ is the input connection weight matrix. ${{\boldsymbol W}_{\textbf{out}}}$ is the output connection weight matrix. ${{\boldsymbol W}_{\textbf{res}}}$ is the internal connection weight matrix.
The feedback loop is divided into ${\boldsymbol N}$ nodes at equal intervals, and the time interval between each node is ${\boldsymbol \theta}$; these virtual nodes are used to replace the nodes in the traditional all-connection-based reservoir. The delay loop records the reservoir state information in the past delay period ${\boldsymbol \tau}$, and the internal state is updated by the nonlinear node after one loop. The internal state of the reservoir is determined by the transient response of the nonlinear device in the current input and the feedback loop of the combination of virtual nodes; these virtual node states are weighted linearly with ${{\boldsymbol W}_{\textbf{out}}}$ to produce the output signal. The only matrix in the network that needs to be trained is ${{\boldsymbol W}_{\textbf{out}}}$, and the rest of the matrices can be randomly generated under certain conditions. Therefore, the state function ${\boldsymbol x}({\boldsymbol n})$ of the network node and output function ${\boldsymbol y}$ can be obtained as shown in Eqs. (1) and (2), where ${\boldsymbol u}({\boldsymbol n})$ is the input signal, and ${\boldsymbol \varphi}$ is the bias, which can make the network work on the optimal nonlinear node:
Therefore, the training of the output connection weight matrix is also easy. During training, ${\boldsymbol y}$ in Eq. (2) is replaced by ${\boldsymbol T}$, so the solution of ${{\boldsymbol W}_{\textbf{out}}}$ is shown in Eq. (3):
Considering that the ${\boldsymbol x}({\boldsymbol n})$ matrix may be a singular matrix, the ridge regression algorithm is used to solve this problem and improve the generalization ability of the network at the same time [28], as shown in Eq. (4), where reg is the regularization parameter, and ${\boldsymbol E}$ is the identity matrix. Finally, ${{\boldsymbol W}_{\textbf{out}}}$ is shown in Eq. (4):
B. Single-Loop Optoelectronic Reservoir Computing
The single-loop optoelectronic RC scheme is shown in Fig. 2. The continuous-wave laser is used as the input of the MZM, and the optical carrier is modulated by the modulation voltage of the MZM. At this time, the MZM is in the working state of intensity modulation. The output of MZM is the state function ${\boldsymbol x}({\boldsymbol n})$, which is collected by a photodetector (PD) and used as the loop output. It is then injected into the delay feedback loop. After the optical attenuator, it is converted into an electrical signal by a photodetector and combined with the preprocessed input radar signal to generate the modulation voltage of the MZM, forming a feedback loop optoelectronic RC scheme.
Fig. 2. Schematic diagram of the optoelectronic RC. LD, laser diode; MZM, Mach–Zehnder modulator; OA, optical attenuator; PD, photodetector; AMP, amplifier; BPF, bandpass filter.
The internal response is shown in Eqs. (5) and (6), where ${\boldsymbol x}({\boldsymbol n})$ is the state function, ${{\boldsymbol \tau}_{\boldsymbol H}}$ is the time constant of the high-pass filtering effect of AMP, ${{\boldsymbol \tau}_{\boldsymbol L}}$ is the time constant of the low-pass filtering effect of PD, and ${\boldsymbol \beta}$ is the normalized feedback coefficient. The function of ${\boldsymbol \alpha}$ is equivalent to the internal connection weight matrix ${{\boldsymbol W}_{\textbf{in}}}$, which corresponds to the modulation of the state function by the optical attenuator, ${\boldsymbol \tau}$ is the delay of the feedback loop, ${\boldsymbol \gamma}$ is the scaling parameter for the input signal, ${\boldsymbol \varphi}$ is the bias voltage of MZM, and ${{\boldsymbol V}_{\boldsymbol \pi}}$ is the half-wave voltage of the modulator. The typical values of the fixed parameters in the system are presented in Table 1 [29]. The other parameters are optimized in the numerical study to optimize the human target recognition performance.
Table 1. Typical Values of Some Parameters in the System
In this paper, we also introduce a parallel dual-loop optoelectronic RC scheme with a dual-polarization Mach–Zehnder modulator (DPol-MZM) for human target recognition (details can be found in our previous work in [24]):
C. Input Radar Signal Preprocessing
The radar signals used for analysis are obtained from the public IDRad data set [12]. Radar signals for this data set were collected using a 77 GHz multichannel frequency-modulated continuous wave (FMCW) radar platform. According to [12], the SNR of the data provided by the radar signal varies from 10 dB for targets within a range of 1 m to 7 dB for targets located at around 8 m, which is also one of the challenges faced by the human recognition task. This data set consists of three parts: training set, test set, and validation set. The radar signal recorded in the training set consists of two parts. The first part records the random movement data of five people in the same room alone for five consecutive minutes, including turning, stationary, etc. The postures of the five people are shown in Table 2. The second part is the second-round recording after two weeks. The difference between these two times is that five people wear different clothes and shoes, and the duration is 15 min per person. The test and validation sets were recorded in another room with more reflections and disturbances, and the duration was 5 min per person. Therefore, this data set fully simulates realistic human recognition scenarios but, at the same time, makes this task extremely challenging. Moreover, each second contains about 15 frames. We perform a 2D Fourier transform on the signal of each frame of the data set to obtain a distance-Doppler map, sum it in the distance dimension to obtain a micro-Doppler feature with a 1 $*$ 256 column vector for each frame signal, and then use this vector as the input signal of RC. Figure 3 shows a micro-Doppler feature map of a roughly 17 s segment.
Table 2. Physical Characteristics of the Targets in the IDRad Data Set [12]
3. RESULTS AND DISCUSSION
Although the multiple connection matrices of RC are randomly generated, several parameters can still directly influence the performance of RC in radar signal processing. In this article, we optimize some essential parameters. Here, we used frame error rate (FER) to represent the performance of RC for human recognition based on radar signals. In addition, the process of optimization is carried out under the condition that the training and testing of the network use the first 5 min and the last 15 min of the training set, respectively; that is, the training and testing data are collected from the same room.
A. Input Bundled Frames
Even if the micro-Doppler features are extracted from the radar echo signal, the complexity of the signal is still high. Therefore, if each frame is randomly scrambled for training and testing, the contextual correlation of the input signal is too weak for the network to have good recognition performance. To this end, the solution we take is to bundle multiple frames of the same target and shuffle them randomly. As shown in Fig. 4, the number of bundled frames corresponds to the horizontal axis. The number of bundled frames should be as short as possible in practical applications; otherwise, it will take longer segments to identify the targets. When the input length increases from 1 to 100, the recognition accuracy shows good improvement; when it is further increased, there is no obvious improvement. Therefore, the input length is selected as 100 frames, that is, a signal fragment of approximately 6.5 s.
B. Size of Reservoir and Regularization
The number of virtual nodes ${\boldsymbol N}$ is the size of the reservoir, which is an important parameter affecting the performance of RC. In Eq. (3), the size of ${\boldsymbol N}$ affects the number of solutions to the output connection weight matrix ${{\boldsymbol W}_{\textbf{out}}}$. When ${\boldsymbol N}$ is too small, the matrix equation is almost insoluble, so the obtained ${{\boldsymbol W}_{\textbf{out}}}$ may not even fit the data of the training set, let alone the new sample data, and the network is in a state of under-fitting. When ${\boldsymbol N}$ is too large, the solution of the matrix equation will accurately fit the data of the training set, but it lacks the generalization ability and cannot adapt to new samples; the network is also in a state of overfitting. The regularization parameter reg can adjust the generalization ability of the network. Therefore, the optimal values of these two parameters are affected by each other.
As shown in Fig. 5, only when ${\boldsymbol N}$ is large will there be a relatively low error rate. At this time, the optimal reg will also be relatively large. When the optimal reg corresponding to ${\boldsymbol N}$ is selected, we can obtain the relationship between ${\boldsymbol N}$ and the recognition accuracy in Fig. 6.
It can be roughly seen that, when ${\boldsymbol N}$ increases from 50 to 500, the FER of target recognition decreases rapidly, and the performance gain brought by increasing the size of the reservoir is obvious. When it is greater than 500, the performance improvement brought about by the increase of the reservoir node ${\boldsymbol N}$ starts to become slow. When ${\boldsymbol N} = 2000$, FER can be reduced to about 13%, but its training and testing time is about six times longer than when ${\boldsymbol N} = 1000$. Considering the calculation amount and performance benefits of the system, in the subsequent optimization process, the selections of the two parameters of ${\boldsymbol N}$ and reg are 1000 and ${2} \times {10\rm E \text{-} 10}$, respectively. When ${\boldsymbol N} = 1000$, the training data equal to 60,000 frames, and the test data equal to 20,000 frames.
C. Normalized Feedback Coefficient
The size of the feedback ${\boldsymbol \beta}$ is also an important parameter that affects the performance of RC. The relationship between the FER and the normalized feedback coefficient ${\boldsymbol \beta}$ is shown in Fig. 7, which is obtained by averaging after two-round simulations. It can be seen that the lowest FER is achieved when ${\boldsymbol \beta}$ is between 0.1 and 0.7. When ${\boldsymbol \beta} = 0$, the network loses its time memory ability; if ${\boldsymbol \beta}$ is too large, it may lead to the divergence of the network state.
D. Internal Connection Weight Matrix
Figure 8 shows the performance with the adjustment of the spectral radius of the internal connection weight matrix, which is the feedback gain. Too large or too small value will lead to poor training due to improper weighting of ${\boldsymbol x}({\boldsymbol n} - {\boldsymbol \tau})$. From the fitted curves, we can see that the general trend is that the FER decreases and then increases as ${\boldsymbol \alpha}$ increases and is low when ${\boldsymbol \alpha}$ is set between 0.4 and 0.6. Thus, we set ${\boldsymbol \alpha}$ in the matrix to 0.55, so the spectral radius is also equal to 0.55.
E. Final Recognition Results and Comparison
To better verify the performance of the RC on the task of human recognition, two kinds of optoelectronic RC are used in our numerical study work. One is single-loop optoelectronic RC; the other is DPol-MZM-based optoelectronic RC. The DPol-MZM-based optoelectronic RC has been verified to have stronger performance in dealing with nonlinear signals by utilizing the polarization of light [24]. The system parameter settings after simulation optimization are shown in Table 3. The size of the single-loop RC is the sum of the number of nodes set by the two feedback loops of the dual-loop RC.
Table 3. Single- and Dual-Loop Optoelectronic RC Parameter Settings
As we can see from the table above, the delay of the feedback loop is about 1E-7 s after optimization. If this parameter is too small, it will put higher requirements on the sampling frequency of the input and output devices; if it is too large, the internal response of the reservoir will tend to be stable, and performance will be reduced. It takes a delay period to process a frame of micro-Doppler signal, so theoretically, the reservoir part only takes 8 ms to process 80,000 frames. This also reflects the high efficiency of using the hardware optoelectronic RC scheme to handle this task. This calculation does not include the part of computer processing, such as the linear operation of the state function ${\boldsymbol x}({\boldsymbol n})$, the calculation of the output connection weight matrix ${{\boldsymbol W}_{\textbf{out}}}$, etc.
We show part of the optoelectronic RC output for classifying each target in Fig. 9. This is the case when the training data set and test data set are collected from the same room. Since the optoelectronic RC is memorized, the previous input segment will affect the current output, and this will affect the performance of RC in classification. It can be seen from the second and fifth segments that the classification results are comparatively confusing.
Here, we considered the performance of the single- and dual-loop RC with training and testing data from the same room or different rooms.
Table 4. Target Recognition Accuracy in Four Cases
Table 4 shows the classification of five targets in four cases. Table 4(a) and 4(c) are the cases where both training and testing data are from signals collected in the same room of the training set, corresponding to the results of single-loop RC and dual-loop RC, respectively. Table 4(b) and 4(d) are the cases where training and testing data are from signals collected in different rooms, corresponding to the results of the single-loop RC and dual-loop RC, respectively. The target classification of the four cases is obtained by averaging 10 numerical results. It can be seen that the accuracy is higher when training and testing are in the same room, and the error is mainly caused by the judgment of the first target. From Table 2, the size of the first and third targets is quite close, so many frames of the first target are misjudged as the third target. The second, third, fourth, and fifth targets are judged with higher accuracy in the same scene. On the premise that the target can walk randomly indoors, RC shows good performance.
Human recognition in different rooms is more challenging; at the same time, however, the accuracy is also lower due to the increase of interference factors. It can be seen in Tables 4(b) and 4(d) that the most serious misjudgments are the second and fifth targets. The second target was more misjudged than the third. Meanwhile, the fifth target was more misjudged than the first target. In this case, in addition to body posture, the error factors are caused by noise factors such as environmental differences and environmental interference, so we cannot simply be considered the interference term of posture.
Consider different RC system performances. In the same room, compared with the single-loop RC, the accuracy of the judgment of the first target is improved by about 5.5%, while the judgment of the other four targets is almost the same, and it can better distinguish the first and third targets. When the training and test data come from different rooms, the dual-loop RC also improves the judgment of the first target, which is about 12% higher than that of the single-loop RC, but the judgment accuracy of the second target drops by about 2%.
We presented the overall FER of the five targets for 10 simulations under four conditions, as shown in Table 5. It can be seen that the maximum and minimum values of the simulated data in the four cases do not deviate significantly from the average, so the data are not discarded; the performance of the dual-loop RC is reduced by 1.2% when the training set and the test set come from the same room; when the data come from different rooms, the error is reduced by about 2%, reaching 40%. This shows that the dual-loop optoelectronic RC does not improve significantly when processing high-complexity signals, which is different from its processing of nonlinear signals.
Table 5. Overall Recognition Performance of FER in Four Cases
In the optimization process and the comparison of the two kinds of optoelectronic RC, to reduce the network running time, we choose the size of the reservoir as 1000. But, compared with other tools, a better $N = {2000}$ was selected. Other tools here refer to principal component analysis (PCA) in combination with a support vector machine (SVM) and a random forest (RF) classifier, deep convolutional neural networks (DCNN), DCNN based on initial block (IB), and structured inference network (SIN) plus long short-term memory (LSTM) [14]. Table 6 compares the performance of the six tools when training and testing originate from different rooms.
Table 6. Recognition FER with Six Tools
The performance of the optoelectronic RC is comparable with that of the SVM and RF, but it lags behind the methods based on the DCNN and the LSTM. However, the other methods are much more complex, such as the SVM, the convergence is slow when processing the multidimensional input. The time complexity of the CNN needs to be accumulated within and between layers, and the time complexity of each convolutional layer is obtained by multiplying the output feature map area, the convolution kernel area, and the number of input and output channels. Moreover, the DCNN is more complex since it combines multilayer deep neural connections. The LSTM is a deep learning method of the recurrent neural network, and the calculation of the RC is a simplified network of the recurrent neural network, so the calculation of the RC is greatly reduced in complexity. Furthermore, there are rarely direct hardware implementation reports for the other methods, while the hardware implementation scheme of the optoelectronic RC is relatively simple and feasible, which can take advantage of the fast processing speed and low power consumption in the optical domain.
4. CONCLUSIONS
In this paper, we explore the performance of optoelectronic RC schemes on the task of human target recognition based on radar signals as close to reality as possible. Through numerical study, we optimized and analyzed the parameters of the single-loop RC system and the dual-loop RC system to better verify the performance of the RC; we also compared the optimal results of the two RC schemes. In the scenes of different rooms and the same room, the accuracy of the judgment of the five targets reached about 63.2% and 85%, respectively. Compared with the single-loop RC, the performance of the dual-loop is not much improved. This suggests that the structure of the dual-loop RC may be limited in application to this task. Although its performance lags behind DCNN, the efficient processing speed of optoelectronic RC still shows that it is a powerful tool in the field of human recognition. The hardware solution of optoelectronic RC can also take advantage of light in practical applications and give human target recognition the possibility of real-time processing. Therefore, the optoelectronic RC scheme is considered as a potential tool for human target recognition tasks in the future.
Funding
National Key Research and Development Program of China (2020YFB1805700); National Natural Science Foundation of China (62101483); Natural Science Foundation of Zhejiang Province (LQ21F010015); Zhejiang Lab (2020LC0AD01); Vetenskapsrådet (2022-04798).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
REFERENCES
1. S. Preussler, F. Schwartau, J. Schoebel, and T. Schneider, “Photonically synchronized large aperture radar for autonomous driving,” Opt. Express 27, 1199–1207 (2019). [CrossRef]
2. X. Gao, S. Roy, and G. Xing, “MIMO-SAR: a hierarchical high-resolution imaging algorithm for mmwave FMCW Radar in autonomous driving,” IEEE Trans. Veh. Technol. 70, 7322–7334 (2021). [CrossRef]
3. F. Amato, A. Farina, M. Fiorini, and S. Gallone, “Surveillance unattended foliage penetrating radar for border control and homeland protection,” TransNav, Int. J. Mar. Navig. Saf. Sea Transp. 7, 193–196 (2013). [CrossRef]
4. L. B. Liu, M. Liu, and J. Q. Wang, “Electromagnetic environment comprehension for radar detection of vital signs at China National Training Center for earthquake search & rescue,” 16th International Conference on Ground Penetrating Radar (GPR), Hong Kong, China, 2016, pp. 1–4.
5. J. Kylmälä, V. Salminen, A. Tuohimaa, and T. Lensu, “Search radar modification for long range ISAR target recognition,” 14th International Radar Symposium (IRS), Dresden, Germany, 2013, pp. 597–600.
6. V. C. Chen, F. Li, S.-S. Ho, and H. Wechsler, “Micro-Doppler effect in radar: phenomenon, model, and simulation study,” IEEE Trans. Aerosp. Electron. Syst. 42, 2–21 (2006). [CrossRef]
7. G. Garreau, C. M. Andreou, A. G. Andreou, J. Georgiou, S. Dura-Bernal, T. Wennekers, and S. Denham, “Gait-based person and gender recognition using micro-Doppler signatures,” IEEE Biomedical Circuits and Systems Conference (BioCAS), San Diego, California, 2011, pp. 444–447.
8. D. Tahmoush and J. Silvious, “Radar micro-Doppler for long range front-view gait recognition,” IEEE 3rd International Conference on Biometrics: Theory, Application, and System, Washington, DC, 2009, pp. 1–6.
9. X. Qiao, T. Shan, and R. Tao, “Human recognition based on radar micro-Doppler signatures separation,” Electron. Lett. 56, 195–196 (2020). [CrossRef]
10. Z. Ni and B. Huang, “Gait-based person identification and intruder detection using mm-wave sensing in multi-person scenario,” IEEE Sens. J. 22, 9713–9723 (2022). [CrossRef]
11. J. Wu, J. Wang, Q. Gao, M. Pan, and H. Zhang, “Path-independent device-free gait recognition using mmwave signals,” IEEE Trans. Veh. Technol. 70, 11582–11592 (2021). [CrossRef]
12. B. Vandersmissen, N. Knudde, A. Jalalvand, I. Couckuyt, A. Bourdoux, W. D. Neve, and T. Dhaene, “Indoor person recognition using a low-power FMCW radar,” IEEE Trans. Geosci. Remote Sens. 56, 3941–3952 (2018). [CrossRef]
13. J. Pegoraro, F. Meneghello, and M. Rossi, “Multiperson continuous tracking and identification from mm-wave micro-Doppler signatures,” IEEE Trans. Geosci. Remote Sens. 59, 2994–3009 (2021). [CrossRef]
14. Z. Li, J. Le Kernec, F. Fioranelli, O. Romain, L. Zhang, and S. Yang, “An LSTM approach to short-range personnel recognition using radar signals,” IEEE Radar Conference (RadarConf), Atlanta, Georgia, 2021, pp. 1–6.
15. Z. Xia, G. Ding, H. Wang, and F. Xu, “Person identification with millimeter-wave radar in realistic smart home scenarios,” IEEE Geosci. Remote Sens. Lett. 19, 3509405 (2022). [CrossRef]
16. H. Jaeger and H. Haas, “Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication,” Science 304, 78–80 (2004). [CrossRef]
17. Z. Shi and M. Han, “Support vector echo-state machine for chaotic time series prediction,” IEEE Trans. Neural Netw. 18, 359–372 (2007). [CrossRef]
18. S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Comput. 9, 1735–1780 (1997). [CrossRef]
19. A. Argyris, J. Bueno, and I. Fischer, “Photonic machine learning implementation for signal recovery in optical communications,” Sci. Rep. 8, 8487 (2018). [CrossRef]
20. Editorial, “Big data needs a hardware revolution,” Nature 554, 145–146 (2018). [CrossRef]
21. P. Antonik, F. Duport, M. Hermans, A. Smerieri, M. Haelterman, and S. Massar, “Online training of an opto-electronic reservoir computer applied to real-time channel equalization,” IEEE Trans. Neural Netw. Learn. Syst. 28, 2686–2698 (2017). [CrossRef]
22. L. Appeltant, M. C. Soriano, G. Van der Sande, J. Danckaert, S. Massar, J. Dambre, B. Schrauwen, C. R. Mirasso, and I. Fischer, “Information processing using a single dynamical node as complex system,” Nat. Commun. 2, 468 (2011). [CrossRef]
23. L. Larger, M. C. Soriano, D. Brunner, L. Appeltant, J. M. Gutierrez, L. Pesquera, C. R. Mirasso, and I. Fischer, “Photonic information processing beyond Turing: an optoelectronic implementation of reservoir computing,” Opt. Express 20, 3241–3249 (2012). [CrossRef]
24. X. Feng, L. Zhang, X. Pang, and X. Yu, “Numerical study of parallel optoelectronic reservoir computing to enhance nonlinear channel equalization,” Photonics 8, 406 (2021). [CrossRef]
25. S. Cai, M. Wang, M. Han, B. Wu, J. Sun, and J. Zhang, “Enhanced performance of a reservoir computing system based on a dual-loop optoelectronic oscillator,” Appl. Opt. 61, 3473–3479 (2022). [CrossRef]
26. D. Brunner, M. C. Soriano, C. R. Mirasso, and I. Fischer, “Parallel photonic information processing at gigabyte per second data rates using transient states,” Nat. Commun. 4, 1364 (2013). [CrossRef]
27. F. Duport, B. Schneider, A. Smerieri, M. Haelterman, and S. Massar, “All-optical reservoir computing,” Opt. Express 20, 22783–22795 (2012). [CrossRef]
28. A. Hoerl and R. Kennard, “Ridge regression: Biased estimation for nonorthogonal problems,” Technometrics 42, 80–86 (2000). [CrossRef]
29. Y. Chen, L. Yi, J. Ke, Z. Yang, Y. Yang, L. Huang, Q. Zhuge, and W. Hu, “Reservoir computing system with double optoelectronic feedback loops,” Opt. Express 27, 27431–27440 (2019). [CrossRef]
