I. Introduction
Target detection using radar in low signal-to-noise ratio (SNR) environments has been a topic of research interest. Improved detection performance extends the range at which targets can be identified. This range is particularly important as the prevalence of low radar cross-section (RCS) targets continues to grow, making it crucial to monitor their activities. For example, drones present potential risks in densely populated areas, such as parks, schools, stadiums, and urban streets. Early detection and tracking of drones at long distances are crucial for providing timely warnings and enabling effective control. Therefore, to maximize drone detection capabilities, it is essential to make target detection algorithms efficient.
Since target detection is a basic yet critical function of radar, many algorithms have been developed to improve radar detection rates. One of the most common types used in target detection is the constant false alarm rate (CFAR) algorithm. A CFAR algorithm measures signal strength by evaluating clutter and noise levels in nearby cells and uses this information to distinguish between targets and noise. Because CFAR is adaptable to cluttered environments, it is widely employed in diverse radar applications [1]. However, in recent years, considerable research effort has been directed toward developing new radar detection algorithms that improve radar performance and overcome existing limitations. In [2], the proposed CFAR algorithm facilitates target detection by exploiting the correlation between linear measurements of radar intermediate frequency signals and sensing matrices in frequency-modulated continuous wave (FMCW) radar. In [3], radar images obtained from multiple-input multiple-output radar are compressed into one or more dimensions to determine the threshold of the proposed CFAR algorithm. These papers propose new threshold determination methods to complement the weaknesses of traditional CFAR algorithms.
In a separate stream of research, deep learning algorithms have been introduced to enhance target detection probability. Combinations of deep neural networks (DNNs) and CFAR algorithms have been developed, with the DNNs achieving the best detection rates [4–6]. Deep learning algorithms integrating convolutional neural networks (CNNs) [7–11], neural networks [12], and U-Net [13–15] provide more accurate and efficient signal processing and object detection capabilities.
Additionally, the use of autoencoders for target detection has enabled efficient performance [16–19]. Autoencoders are effective in learning low-dimensional representations of data, thereby enabling noise reduction and signal enhancement. Although methods have been proposed to increase SNRs by using autoencoders, there is still room for further improvement through the integration of various autoencoder models and training methods. Furthermore, to the best of our knowledge, the use of autoencoders to improve radar range resolution and Doppler resolution has not been investigated.
In this paper, we propose a CNN-based denoising autoencoder to improve target detection probabilities as well as ranges and Doppler resolutions in range-Doppler diagrams. A denoising autoencoder is a variation of the basic autoencoder, primarily used in deep learning to remove noise from data and reconstruct signals. The use of denoising autoencoders has been demonstrated to effectively remove noise from radar images in contexts involving time-frequency images [20], Doppler-enhanced frontal images [21], and remote sensing images [22]. Building on these studies, we aim to apply a denoising autoencoder to range-Doppler diagrams to remove noise from FMCW radar and perform low-RCS target detection. In this study, the denoising autoencoder was trained to generate noise-free range-Doppler diagrams with improved SNRs from noisy range-Doppler diagrams. Instead of training conventional autoencoders, we used noise-contaminated radar images as input data and noiseless radar images as output data in the training process. When this training strategy is adopted, the denoising autoencoder not only enhances the target’s signal components and reduces noise levels but also attenuates the target’s sidelobe levels, thereby improving the target detection resolution. The performance of such an autoencoder has not previously been compared with conventional detection methods, and these strategies have not yet been fully explored. Therefore, by proposing a denoising autoencoder along with our novel training strategy, this research makes several important contributions: (1) improving ranges and Doppler resolution in target detection through sidelobe reduction, (2) improving target detection rates by enhancing targets’ SNRs through noise reduction, and (3) comparing the proposed autoencoder method with CFAR algorithms using the receiver operating characteristics (ROC) curve
To verify our algorithm, we measured drones. Drones are low-RCS targets, so they are difficult to detect when the noise level is high. Specifically, we verified the performance of the proposed algorithm through dual drone measurements. Finally, the trained denoising autoencoder was evaluated by comparing its performance with that of two-dimensional (2D) CFAR algorithms [23] in terms of ROC curves and computational complexity.
II. Training the Denoising Autoencoder with Range-Doppler Diagrams
An autoencoder is one of the most common unsupervised learning tools used for feature extraction in deep learning. The autoencoder compresses the input data into a low-dimensional latent representation and then attempts to reconstruct it into a form as close as possible to the original data. Therefore, the output of the decoder is designed to be the same as the input to the encoder. To achieve this symmetry, the neural network employed primarily comprises an encoder and a decoder. The encoder maps the input data into a low-dimensional feature vector that contains the essential representation of the data. The decoder uses the feature vector generated by the encoder to reconstruct the original data.
A denoising autoencoder is a specialized type of autoencoder designed for noise removal. Unlike conventional autoencoders, this autoencoder is based on a CNN. It incorporates CNN layers in both the encoder and the decoder. The CNN is used to detect spatial features in image data, enabling the denoising autoencoder to effectively capture the spatial structure and characteristics of the input data—except the noise. We propose the use of a denoising autoencoder to reconstruct range-Doppler diagrams with noise removal capabilities, thereby enhancing target detection capacity and improving resolution. The most important function of the denoising autoencoder is to properly structure data from input to output. In this study, the autoencoder was trained to reconstruct noisy images (input data) into noiseless images (output data).
1. Generation of FMCW Simulation Data
To train the autoencoder, we used simulated range-Doppler diagrams. FMCW radar generates periodic chirp signals. The transmitted frequency for one chirp can be expressed as
where fc is the carrier frequency, B is the operating bandwidth, and Tc is the chirp duration. The received signal is delayed in accordance with the distance to the target. The transmitted and received signals pass through a mixer and a low-pass filter (LPF) to produce a beat signal. Finally, the output of the LPF is sampled by the analog-to-digital converter (ADC). Assuming there are L targets, the resulting signal after ADC sampling can be expressed as
where Rl is the range of the l-th target, fd is the Doppler frequency, c is the speed of light, fs is the sampling frequency, n is the sample index of one chirp, and p is the index of the chirp. If vl is the relative velocity of the l-th target, then the Doppler frequency becomes fd = 2vl · fc/c.
After ADC sampling, the signal takes the form of a 2D matrix. In (2), the target’s range information can be estimated by applying a fast Fourier transform (FFT) to the sampling axis (fast time), and the Doppler frequency can be estimated by applying the FFT to the chirp axis (slow time). Therefore, the range-Doppler diagram can be obtained all at once through 2D FFT. The process for generating the range-Doppler diagram is shown in Fig. 1.
Given the location and velocity of the target, we can simulate range-Doppler diagrams while controlling the SNRs of the targets and the Gaussian noise level in the simulation process. When performing the fast Fourier transformation to construct the range-Doppler diagrams, Hamming windows were applied to both the range domain and the Doppler domain, where they served as low side-lobe taper windows.
2. Training the Autoencoder using Simulation Data
Based on the FMCW radar equation, simulation data were generated using the targets’ positions, velocities, and RCS as input. During the simulation, the precise position and velocity of the targets were used as ground truth information. This ground truth represents the target coordinates as a [1,1] pixel matrix composed of values of 1 and 0.1 under the assumption of a state without noise or sidelobes. Noise-contaminated range-Doppler diagrams were used as the input data, while the ground truth— the range-Doppler diagram without noise or sidelobes—served as the output data. By applying a CNN-based denoising autoencoder to the range-Doppler diagram, we can learn the features and patterns inherent in the data while removing the noise.
The denoising autoencoder used in the study features the architecture depicted in Fig. 2. The encoder section processes the input data using convolutional layers with the ReLU activation function. The padding is set to “same,” the strides are set to 2, and the kernel size is set to 4. Setting the padding to “same” ensures that the information at the edges is preserved without loss while maintaining the same size for both the input and output. The encoder consists of two layers, with 32 filters in the first layer and 16 in the second layer. The decoded section is designed symmetrically with the encoder section. Convolutional transpose layers are used to decode feature vectors. A convolutional layer with 1 filter is employed in the last part of the decoder to generate the output image. To determine the optimal structure for the denoising autoencoder, we tested its performance on 10,000 test data samples by varying the number of layers and kernel size. The probability of false alarm (PFA) for the denoising autoencoder was set at 1.0 × 10−7. Fig. 3 presents the accuracy evaluation according to the denoising autoencoder’s structure, confirming that the structure presented in Fig. 2 is the optimal match for the simulation dataset.
To generate the simulation data, radar parameters were set as follows: the center frequency was 77 GHz, the sampling bandwidth was 76.8 MHz, the chirp duration was 6.57 μs, and there were 256 time samples per chirp. Two targets were simulated, and each target was assigned a random range of either 75 m or 135 m and a random Doppler value between 20 m/s and 75 m/s. The RCS values of the targets were also set randomly.
A range-Doppler diagram was generated using a 2D FFT. Fig. 4(a) and 4(b) represent simulated data for two drone targets; these data were used as input for the autoencoder. The input images were created as 64 × 64 arrays. The powers of each pixel were normalized between 0 for the lowest noise value and 1 for the highest target value in the diagram. Fig. 4(c) and 4(d) represent the generated output data, which represent an image in which all the noise has been removed, except at the target location. The precise coordinates of the two targets were identified and marked as array values of 1. The targets were set to a size of 1 × 1, representing the exact coordinates of the targets without sidelobes. Values other than the target coordinates in Fig. 4(c) and 4(d) were set to 0.1. It was heuristically determined that setting them to 0.1 facilitated better training results. The training data consisted of 117,050 pairs of images; 80% of the dataset was allocated for training, while 20% was reserved for the validation set.
To train the denoising autoencoder, the mean squared error loss function with the Adam optimizer was used. The training process was set to run for 10 epochs with the option to shuffle the data. The resulting trained model achieved a train loss of 4.67 × 10−5 and a validation loss of 5.75 × 10−5.
III. Autoencoder Results Based on Simulation Data
After training the denoising autoencoder, we input the simulated test data (Fig. 5) and examined the corresponding output (Fig. 6). The results indicate that the high-intensity points of the signal were accurately reconstructed at the target locations in the output. The SNR was calculated using Eq. (3):
where Psignal refers to the target signal power, while Pnoise refers to the average noise power, excluding the signal. While the SNR is 12.76 dB in Fig. 5(d), it becomes 40 dB in Fig. 6(d), which is the lowest among the Fig. 5 samples. This indicates that we successfully improved the target SNRs using the proposed denoising autoencoder.
Fig. 7 presents simulated test data consisting of two closely positioned targets. Fig. 7(a) shows the case in which the targets have similar velocities but slightly different distances, while Fig. 7(b) shows the converse case, where the distance is the same but the velocities differ slightly. In both figures, the target is placed with a separation of 1 × 1 pixel in the image array. Given the radar parameter used in these simulations, the range resolution is 1.95 m, and the Doppler resolution is 1.17 m/s.
The denoising autoencoder was also trained using range-Doppler diagrams with the same range and Doppler resolution. Fig. 7(c) and 7(d) display the autoencoder outputs for the corresponding test data in Fig. 7(a) and 7(b). The sidelobes and noises were effectively removed, allowing both targets to be adequately detected, which was not possible using the CFAR or conventional autoencoder. In addition, the SNR and signal-to-sidelobe ratio (SSR) increased at the precise coordinates of the targets. SSR was calculated using Eq. (4):
IV. Performance Comparison with CFAR
1. ROC Curve Comparison
To evaluate the performance of the denoising autoencoder, we compared it against three CFAR algorithms: cell averaging (CA-CFAR), greatest of cell averaging (GOCA-CFAR), and smallest of cell averaging (SOCA-CFAR). As the logarithmic scale was used in the simulated range-Doppler diagrams, we excluded order statistic CFAR algorithms because it has been reported that they do not function properly when the logarithmic scale is used. The guard band size of the CFAR algorithms was set as 2 × 2, while the training band size was set as 4 × 4. We converted the test data to a linear scale and compared the performance of the CFAR algorithms and the trained denoising autoencoder using ROC curves.
For this process, we had to integrate post-processing because there was a slight difference in the types of data output by the trained autoencoder and the CFAR algorithms. The autoencoder output appeared as an array of values ranging from 0 to 1, while the CFAR algorithm produced binary outputs of either 0 or 1. In the denoising autoencoder, a target was identified when the output value from the denoising autoencoder was greater than the fixed threshold. By applying this approach to the test data, the PFA can be calculated according to the threshold. In this study, the minimum PFA value for the denoising autoencoder was 2.4 × 10−8, which means that one false alarm occurs in a test dataset of 10,000 samples, each with a size of 64 × 64. If equal PFAs were calculated using the denoising autoencoder and CFAR algorithms, we established identical conditions for comparing their target detection capabilities. We calculated the detection probability offered by different algorithms for each SNR at the same PFA.
When comparing algorithmic performance, we deemed it important to understand the trends in detection probabilities for various SNRs, so ROC curves were ultimately plotted at the same PFA. To draw the ROC curves, we generated 10,000 test data samples for each of the 15 different SNR values, resulting in a total of 150,000 test samples. Fig. 8 represents the resultant ROC curves for CA-CFAR, GOCA-CFAR, and SOCA-CFAR using the trained denoising autoencoder, corresponding to PFA values of 1.01 × 10−6, 1.04 × 10−7, and 2.4 × 10−8. When the average PFA is low, the accuracy increases but the detection probability decreases, and vice versa.
It can be observed that the autoencoder exhibited enhanced target detection performance compared to the CFARs in all three PFA cases. Traditional CFAR techniques use the difference in signal levels between a target and its surrounding noise to detect the target. However, this method’s reliability can degrade in environments with high or irregular noise. In contrast, the denoising autoencoder operates by learning and capturing the unique features of the target. This grants the denoising autoencoder superior performance in high-noise environments compared to CFAR. We also observed that the denoising autoencoder tends to allow a better detection probability than CFAR even with test data on a dB scale. The denoising autoencoder exhibited the most significant difference in detection capacity relative to CA-CFAR, with the autoencoder’s detection probability measured be 0.1544 higher when the PFA was 2.4 × 10−8. This difference is larger than the detection probability gap between the CNN detector and the CFAR algorithm reported in [5]. In addition, while the autoencoder presented in [10] was able to improve the signal-to-interference-plus-noise ratio (SINR) by 9.23 dB, we achieved an improvement of more than 15 dB in the SNR with the test data.
2. Radar Resolution Improvement
To statistically evaluate the improvement in the radar resolution of the denoising autoencoder, the target detection probability was calculated according to the SNR, the range difference, and the difference in Doppler value between the two targets. Table 1 categorizes the cases. Two targets were divided into three cases, and the detection probability for each case was evaluated. This test was designed to evaluate the performance of the autoencoder when the two targets were separated by the theoretical resolution limits. In Table 1, Case 1 represents the reference situation—two targets with random distances and random velocities—which matches the previous simulation data from Fig. 8. Case 2 describes two targets with a fixed range difference of 1.95 m and identical Doppler frequencies (randomly assigned). Case 3 involves two targets with a fixed velocity difference of 1.17 m/s and identical ranges (randomly assigned). These separations (1.95 m and 1.17 m/s) correspond to the theoretical radar resolution for the range and Doppler axes, respectively. In other words, we tested the target detection probability when the targets were separated by the resolution limits in Cases 2 and 3.
By varying the SNR, 1,000 data samples were generated for each SNR and then input into the denoising autoencoder. To calculate the detection probability, we assumed that the detection was successful if all targets were correctly detected. The denoising autoencoder’s PFA was set to 1.01 × 10−6, and the detection probability for Case 1 was derived from the results in the range from 13.5 dB to 25 dB in Fig. 8(a).
As Table 1 shows, as the SNR decreases from 25 dB to 15.5 dB, Case 2 consistently shows a lower detection probability than Case 1. The largest difference occurs at 18.35 dB, where Case 2’s detection probability is 3.3% lower. Similarly, Case 3 exhibits a lower probability than Case 1, with a maximum difference of 92% at 18.35 dB. This suggests that the denoising autoencoder maintained stable performance, effectively distinguishing two targets at the radar’s resolution limits across various SNRs, whereas conventional CFAR failed to detect two separate targets.
Table 2 presents the results of a statistical evaluation of the resolution improvement achieved by the denoising autoencoder compared to that achieved using CFAR. In this analysis, the SNR was fixed at 19.5 dB. For each case, one parameter—either the range difference or the Doppler difference—was held constant at the theoretical limit, while the other was varied in fixed intervals to generate the range-Doppler diagram. When the Doppler difference was varied, the range difference was fixed at 1.95 m (the theoretical radar range resolution), simulating a scenario in which the targets were separated along the Doppler axis, as shown in Fig. 7(a). Conversely, when the range difference was varied, the Doppler difference was fixed at 1.17 m/s (the theoretical Doppler resolution), simulating a scenario in which the targets were separated along the range axis, as shown in Fig. 7(b).
Target detection was performed on 1,000 data samples generated for each case. Both the denoising autoencoder and CFAR algorithms were used. A true positive was determined when both targets were detected, as illustrated in Fig. 7(c) and 7(d). The PFA for both CFAR and the denoising autoencoder was set at 1.01 × 10−6. These results confirm that the denoising autoencoder improves both range resolution and Doppler resolution compared to CFAR. Even when two targets could not be resolved using CFAR, they could be detected using the de-noising-autoencoder. While previous studies have failed to provide clear evidence of the ability of deep learning-based detectors to overcome radar resolution limits, the results in Table 2 show that the proposed denoising autoencoder effectively increases detection probability, even when two targets are separated below radar resolution limits.
3. Time Complexity Comparison
A detection algorithm for radar systems should be able to operate under extremely strict real-time constraints. Thus, it is important to investigate the computational complexity of the suggested algorithm. Accordingly, we compared the time required by CFAR and by the denoising autoencoder to process data, thereby evaluating the complexity of the algorithms. Fig. 9 represents the operation time of each algorithm according to data size. We used an Intel i9 13900KF computer with DDR5-4800 RAM and RTX 4070 Ti GPU, where “Data Size 64” refers to data of size 64 × 64. For this size, the denoising autoencoder takes 0.005 seconds, while CFAR requires 0.03115 seconds. Consequently, the denoising autoencoder is approximately 6.23 times faster than the CFAR algorithm. At a size of 1024 × 1024, the denoising autoencoder takes 0.158 seconds, while CFAR takes 0.9324 seconds. In this case as well, the time taken for the denoising autoencoder operation is approximately 5.901 times shorter than that of CFAR. As shown in Fig. 9, the time taken by the denoising autoencoder to process a single data point is six times faster on average than the time needed by CFAR.
V. Evaluation Based on Experimental Data
We conducted an experiment to evaluate the performance of our trained denoising autoencoder. Fig. 10 displays the concrete experiment setup. The radar was mounted at a height of 1 m from the ground, and two 27.5 cm × 27.5 cm drones were programed to approach the radar in a clutter-free environment. We used an AWR1243 millimeter-wave FMCW radar system (Texas Instruments) operating at 77 GHz. The following parameters were employed: 128 chirps, a chirp duration of 57 μs, a chirp slope of 70 MHz/μs, 256 time samples per chirp, and a frequency bandwidth of 4 GHz. Using the FMCW radar, we measured the two drones approaching the radar at different speeds.
Fig. 11(a) presents the results for the case in which the two drones were at a similar distance. The radio detected substantial stationary clutter (at a velocity of 0). To address this issue, we captured the idle environment with the radar and identified the locations of the clutter. We then identified the coordinate axis at which the clutter signals occurred, removed the data from these locations, and filled the removed coordinates with linear interpolations of the surrounding pixels.
Fig. 11(b) presents the data after removing all clutter. Fig. 11(c) presents the output of the CA-CFAR values in Fig. 11(b), while Fig. 11(d) shows the output of the denoising autoencoder. In contrast to the CA-CFAR output, the denoising autoencoder effectively eliminated noise and accurately reconstructed the two drones. This result indicates that even when targets are very close together in a range-Doppler diagram, they can be accurately detected. The minimum range and Doppler resolution were also measured as 1 × 1 in the 64 × 64 array in the experimental data, which corresponds to 1.63 m and 2.11 m/s.
Fig. 12(a) shows the case in which the two drones are at different distances. In Fig. 12(b), the drone located at 3.3 m is farther away than the other drone, resulting in a lower SNR. The process represented in Fig. 12 was conducted similarly to that portrayed in Fig. 11. As shown in Fig. 12(c) and 12(d), the denoising autoencoder accurately detected both drones. This is not the case with the CA-CFAR output, which detected only the drone closer to the radar. This result indicates that the denoising autoencoder can detect targets with low SNRs.
To evaluate the performance of the denoising autoencoder, measurement data representing the two drones were compared with simulated ROC curves. The experiment was conducted using an AWR1243 FMCW radar to facilitate a scenario in which two drones approached the radar at different velocities. Since the maximum measurement range of the radar was 5.5 m, the positions of the two drones were analyzed at 1-m intervals. Specifically, data were collected for 100 frames each in the ranges of 4.5–5.5, 3.5–4.5, and 2.5–3.5 m.
If the distance between the two drones increased, the SNR difference became significant, yielding conditions that were unsuitable for calculating detection probabilities. To address this issue, the distance between the two drones was limited to 1 m during the experiment. This ensured that the SNR difference between the two drones remained within 5 dB, and the distance interval was determined empirically. The experimental data were processed by applying the threshold of the denoising autoencoder, which was calculated as PFA = 1.04 × 10−7 following the simulation tests. The results were analyzed and compared with the performance of CA-CFAR, as shown in Fig. 13.
As shown in Fig. 13, the detection probability results obtained by applying the denoising autoencoder and CA-CFAR to the measured test data were compared with the theoretical results. For the experimental data, SNRs were calculated by averaging the SNRs of the target across 100 frames, and the detection probabilities were plotted as ROC curves. The experiment confirmed that the denoising autoencoder maintained higher detection probabilities than CA-CFAR across the overall SNR distribution. Specifically, when the SNR was 14.5 dB in the simulation data, the difference in detection probabilities between the denoising autoencoder and CA-CFAR was 0.0609. However, with an SNR of 13.9 dB, the difference increased to 0.12, demonstrating the superior detection performance of the denoising autoencoder.
VI. Discussion
This section compares our device’s target detection capability to that of state-of-the-art technologies. Following the introduction of a DNN-based peak detector in [4], target detectors based on CNNs were proposed in [5–7], and target detectors using autoencoders were suggested in [8–10]. Unlike our denoising autoencoder, which calculates detection probabilities under a fixed PFA, the CNN-based detectors in [5–7] determine detection probabilities while varying the PFA according to the SNR. Therefore, direct comparison is difficult because the PFA settings differ even at the same SNR.
Specifically, in [5], ROC curves were used to evaluate the performance of a CNN-based detector in calculating detection probabilities for varying SNRs. In [5], the PFA was set differently depending on the SNR, whereas in our method, the PFA remains fixed regardless of the SNR. As a result, differences arise in the ROC curves. Additionally, [5] evaluated performance with a single target. This limits the direct comparability of their ROC curves with ours, as we used multitarget environments and confirmed that the denoising autoencoder can effectively detect multiple targets simultaneously. While the detection probabilities of CNN and CFAR in [5] remained stable, even with an SNR that was 10 dB lower than the SNR used during training, the gap between the detection probabilities of our denoising autoencoder and those of CFAR increased as the SNR decreased. This indicates that the denoising autoencoder can achieve better detection probabilities than CNNs at low SNRs.
Studies evaluating the performance of autoencoder-based detection methods have used different metrics to emphasize specific aspects of performance. For example, [8] focused on spatial relationships and evaluated detection performance using 2D correlation coefficients. In contrast, [9] used signal-to-clutter ratio and peak signal-to-noise ratio, which are related to image quality, for the evaluation. Additionally, [10] emphasized performance in environments with interference and noise, using SINR and structural similarity index to measure image quality. These metrics are primarily used to assess results output as 2D images, making it challenging to directly compare them with our results. However, [10] developed an autoencoder-based detector similar to ours, which extracts only target components, and reported that the autoencoder’s output improved the SINR by 9.23 dB in cluttered images. In contrast, our approach aimed for an SNR improvement of up to 15 dB during training.
Thus, although there is limited scope for direct comparisons with previous studies, we have nevertheless confirmed that a denoising autoencoder can significantly improve target detection performance across various SNRs. Our study differentiates itself from prior research by proposing an approach that overcomes the theoretical limitations of radar resolution in target identification. While existing studies lack specific discussions of radar resolution, our study evaluated detection performance based on target range, Doppler resolution, and SNR variations, demonstrating the possibility of practically surpassing theoretical limits on resolution. Thus, our denoising autoencoder effectively extracts target features during training and enhances target detection performance.
VII. Conclusion
This research proposes a denoising autoencoder for refining range-Doppler diagrams to improve target detection and radar resolution. The denoising autoencoder was trained using noisy range-Doppler diagrams as input and noiseless range-Doppler diagrams as output to remove noise and sidelobes. Once the training was completed, the autoencoder was able to identify the precise coordinates of targets and increase SNRs and SSRs. It can distinguish closely spaced targets in range-Doppler diagrams. To evaluate the effectiveness of the proposed denoising autoencoder, its performance was compared with that of various CFAR algorithms in terms of ROC curves and resolution improvement. The algorithms were applied to range-Doppler diagrams constructed using empirically measured drone data. The autoencoder exhibited stronger performance than the CFAR algorithm, demonstrating its potential as a reliable and accurate target-detection algorithm.
PDF Links
PubReader
ePub Link
Full text via DOI
Download Citation
Print
