Radar is an active sensing modality that exploits the round-trip propagation of electromagnetic waves. By transmitting a known waveform and processing the returned echoes, radar can estimate a target’s range, radial velocity (via Doppler shift), and scattering characteristics. Due to its long sensing range and robustness against illumination and adverse weather conditions, radar has become indispensable in aviation, maritime navigation, meteorological monitoring, and intelligent transportation systems. However, most automotive implementations rely on a forward-looking radar with a limited field of view, which inherently leaves occluded regions unobserved. Such blind zones—commonly encountered at urban intersections, sharp corners, and indoor parking structures—pose a major obstacle to reliable perception and safe fully autonomous driving.

To address this limitation, we propose a non-line-of-sight (NLOS) imaging framework that integrates a frequency-modulated continuous-wave (FMCW) radar with specular reflection propagation paths, as illustrated in Fig. 1. By exploiting corner-induced reflections, the proposed approach aims to infer targets that are hidden from direct line-of-sight sensing.

An FMCW radar emits a linear-chirp signal whose instantaneous frequency is given by [53]:(1)f(t)=f0+(BTchirp)⋅twhere f0 is the start frequency, B the sweep bandwidth, and Tchirp the chirp duration. After reflection from a target, the received signal undergoes a time delay and Doppler shift, yielding [53,54] (2)r(t)=cos⁡(2π(f0t+B2Tchirp⋅t2+Δf⋅t))

The term f0t denotes the carrier component of the transmitted signal, BTchirp⋅t2 represents the linearly increasing frequency sweep, and Δf⋅t corresponds to the frequency offset induced by the target’s radial motion. By analysing these variations in the received echo, the radar estimates range and velocity: the time delay is proportional to the target’s distance, whereas the frequency offset reflects its kinematic state. Accordingly, FMCW radar can concurrently provide accurate range and speed information while exhibiting strong immunity to interference, rendering it well-suited to complex traffic environments [22,25–27,53].

In an FMCW radar, the frequency difference—or beat frequency—that arises after mixing the transmitted (Tx) and received (Rx) signals is [53]:(3)fbeat=BTchirp×Δtwhere B denotes the sweep bandwidth (Hz), Tchirp is the chirp duration (s), and Δt is the round-trip propagation time of the waveform. Because this transit time is governed by the target range D, it can be expressed as [53,54]:(4)Δt=2Dcwith c=(3×108 m/s) representing the speed of light. Substituting (4) into (3) yields a direct relationship between the measured beat frequency and the target distance, thereby enabling precise range estimation from the observed spectral peak.

When simulating radar performance in urban environments, geometric occlusion is often used to determine whether a target is obscured by surrounding structures. Let the radar and target positions in the ground plane be denoted by pr and pt, respectively. A parametric line segment connecting the two points is defined as [55] (5)p(t)=pr+t(pt−pr)where the scalar parameter t continuously maps every location between radar (t=0) and target (t=1). If this ray intersects any building footprint—modelled as polygons or rectangles—for some 0<t<1, the direct line-of-sight (LOS) between radar and target is considered obstructed, producing an occlusion that prevents the radar from detecting the target directly [20,55]. Geometric-occlusion checks of this type are integral to sensor layout studies and signal propagation analyses in smart city planning and autonomous driving research. They enable the quantitative assessment of how obstacle distributions constrain radar visibility, thereby supporting high-fidelity performance simulations in realistic urban street scenes.

In automotive radar sensing, angular estimation is crucial for determining the location of obstacles relative to the ego-vehicle. Two angular descriptors are used: the azimuth angle θ in the horizontal (x−y) plane and the elevation angle in the vertical (z) plane. The global (absolute) azimuth of a target with Cartesian coordinates (xt,yt) relative to the radar at (xr,yr) is obtained from [25,29] (6)θglobal=arctan⁡(yt−yrxt−xr)and is conventionally expressed in either the range [−180∘,180∘] or [0∘,360∘]. The elevation angle (not explicitly required in planar street-level models) describes the target’s vertical offset relative to the radar boresight and is derived analogously from the z−component.

For motion-compensated tracking, it is often more useful to express the target direction in the radar’s own body frame, as a relative azimuth. If the sensor (or host vehicle) is yaw by an angle θyaw with respect to the global reference, the raw difference [25,29] (7)θrelative=θglobal−θyawmust be wrapped in a continuous interval. Applying a modulo operation followed by a shift yield [25,29] (8)θrelative=mod(θglobal−θyaw+180,360)−180so that θrelative∈[−180∘,180∘] satisfies the right-hand coordinate convention and admits unambiguous left/right discrimination.

In practice, angular accuracy is degraded by multipath, clutter, and sensor noise; therefore, Monte Carlo perturbations are frequently injected into simulation models to emulate measurement uncertainty, while adaptive filtering and interference mitigation techniques are deployed in real-time systems to restore precision. Accurate estimation of both azimuth and elevation thus remains crucial for localization, navigation, and object classification in autonomous driving applications.

In radar-based perception, a target is deemed directly detectable when its range D does not exceed the sensor’s maximum instrumented distance (20 m), and its bearing falls within the radar’s predefined field-of-view limits. Otherwise, the object must be assessed through indirect mechanisms such as reflection or diffraction, or may be disregarded altogether. Even under severe visibility degradation—caused by fog or other environmental interference—the geometrically derived distance serves as an “ideal-condition” reference against which simulated perturbations (random error, attenuation, propagation delay, etc.) can be superimposed to approximate real-world measurements. The range itself is evaluated via the Euclidean metric [55]:(9)D=(xt−xr)2+(yt−yr)2where (xr,yr) and (xt,yt) denote the radar and target coordinates, respectively. Owing to its computational simplicity, Eq. (9) facilitates rapid, large-scale assessments of spatial relationships between a radar unit and multiple targets in complex urban streetscapes.

The root-mean-square error (RMSE) is an established metric for quantifying estimation accuracy in radar signal processing; a lower value indicates closer agreement between an algorithm’s output and the ground truth. It is defined as [54] (10)RMSE=1N∑i=1N(xi−x^i)2where xi and x^i denote the ith reference and estimated samples, respectively, and N is the total number of observations. The present study employs RMSE in two complementary ways:

Signal-domain assessment—RMSE is computed for the original echoes, the echoes corrupted by environmental interference, and the echoes after neural-network–based interference suppression, thereby quantifying the improvement in waveform fidelity.

This two-tier evaluation framework rigorously validates the effectiveness and practicality of the interference-mitigation approach under realistic radar-measurement conditions.

The signal-to-noise ratio (SNR) is a fundamental metric that quantifies the proportion of useful signal energy to unwanted interference, and it is routinely employed in radar, telecommunications, and image-processing applications to characterize system performance under varying noise conditions. Expressed in decibels, SNR is defined by [54] (11)SNR=10×log10⁡(PsignalPjam)where Psignal denotes the power of the desired signal in an interference-free scenario, and Pjam represents the power of the noise or jamming component measured at the receiver. A higher SNR indicates that the information-bearing component dominates the received waveform, leading to superior detection or decoding accuracy, whereas a lower SNR implies that noise constitutes a substantial fraction of the total power, thereby degrading the system’s ability to discriminate targets from background clutter.

AlexNet [56] is an eight-layer architecture comprising five convolutional and three fully connected layers, originally devised to enhance feature extraction and classification performance on large-scale datasets such as ImageNet. In the present study, the network is repurposed for one-dimensional radar-waveform regression by substituting every two-dimensional convolutional kernel with a one-dimensional counterpart, while preserving the original convolutional-fully connected (FC) hierarchy. Experimental results indicate that the modified Alex Net markedly suppresses environmental noise, yielding a substantial reduction in the root-mean-square error (RMSE) of noisy radar signals and achieving higher accuracy in range and angle estimation. These findings demonstrate the architecture’s strong transferability and practical value in cross-domain signal-processing applications.

To enhance around-the-corner radar perception at urban intersections under heterogeneous environmental interference, we propose an improved AlexNet-derived network for radar-echo restoration, as illustrated in Fig. 2. In noisy intersection conditions, a conventional AlexNet baseline may exhibit limited interference tolerance and unstable outputs, which can propagate into range and azimuth estimation errors. We therefore adopt a minimal-modification strategy that preserves the overall AlexNet backbone while improving feature refinement and training stability.

Specifically, residual learning is introduced by inserting a lightweight residual module after the second convolutional layer (conv2). The module learns a residual mapping via two sequential convolutional layers with batch normalization and rectified linear unit (ReLU) activations. Its output is combined with the conv2 features through an identity skip connection (element-wise addition), followed by a post-activation. Batch normalization is also applied to early features to stabilize optimization under varying interference intensities. The remainder of the network (conv3–conv5 and fc6–fc8) follows the AlexNet-style hierarchy, and the final regression output directly produces a restored radar-echo signal of the same length as the input, without introducing additional decoding stages.

This residual-enhanced design facilitates gradient propagation, mitigates degradation effects, and preserves fine-grained echo characteristics that are critical for accurate range and azimuth-related estimation in urban intersection scenarios. As summarized in Table 1, the proposed architecture differs from the baseline primarily by the addition of a residual module and normalization layers, yielding improved robustness and generalization while maintaining a lightweight structure suitable for real-time deployment.

A MATLAB-based simulation framework was developed to investigate non-line-of-sight (NLOS) blind-spot detection at urban intersections. The virtual environment models a representative intersection geometry with buildings and static obstacles that obstruct direct line-of-sight (LOS) visibility. This setup enables systematic analysis of how structural occlusions degrade radar perception and whether specular reflections from building façades can be exploited to infer targets hidden around corners. Within this framework, the radar processing chain is used to estimate the relative range and bearing of a target with respect to the host vehicle. To quantify the benefit of multipath exploitation, we conduct a comparative evaluation between (i) direct-path (LOS) radar detection and (ii) reflection-assisted (NLOS) detection. Performance is assessed in terms of estimation accuracy and robustness under different occlusion conditions and interference levels.

Conventional automotive radar sensors are commonly mounted near the longitudinal centerline of the vehicle’s front fascia and are primarily oriented forward. While this configuration is effective for unobstructed frontal scenarios, it is inherently limited in intersection corner cases where buildings or parked vehicles create large occluded regions. To mitigate line-of-sight blind spots, we employ an enhanced installation scheme utilizing a lateral array of three 77-GHz millimeter-wave radars co-located on the front bumper. Each radar is mechanically aligned with the vehicle heading, while its effective sensing sector is electronically defined through beam steering. The key parameters are set to a maximum unambiguous range of 20 m and a horizontal field of view (HFOV) of 60 degrees. Specifically, the boresights of the three electronic beams are steered to −60°, 0°, and +60° relative to the vehicle heading. Accordingly, the three radars cover the azimuth intervals [−90°, −30°], [−30°, +30°], and [+30°, +90°], respectively. Concatenation of these sectors yields a continuous 180° forward coverage without gaps, forming a wider observation aperture that is better suited to capturing reflection-induced returns in urban intersection scenarios. This configuration substantially reduces blind zones while maintaining straightforward hardware integration and installation. In addition, the proposed architecture is readily extendable to advanced sensing strategies, such as adaptive beam scanning and multi-sensor fusion. These extensions can support dynamic field-of-view expansion and environment-specific calibration to further improve perception robustness in complex intersection scenarios, as illustrated in Fig. 3.

To examine reflection characteristics and Doppler signatures of millimeter-wave radar in an urban environment with building-induced occlusions, we constructed a representative driving scenario using a commercial simulation platform. The simulated scene provides synthetic radar returns for systematic evaluation, and the generated dataset is subsequently used to develop and train a deep-learning-based radar interference mitigation module, adapted from AlexNet. As illustrated in Fig. 4, the virtual environment represents a four-way intersection formed by two orthogonal, bidirectional roadways with explicit lane markings. The ego vehicle (length: 4.5 m, width: 1.8 m) is initialized 10 m south of the intersection center and oriented northward (yaw = 90°). Three targets are introduced to create both line-of-sight and occluded cases: (i) a bicycle located on the west side of the intersection at (−15 m, −3 m); (ii) a passenger car directly ahead of the ego vehicle at (2 m, 5 m); and (iii) a pedestrian approaching from the northwest at (14 m, 3 m).

Building Reflection Detection Mode refers to a signal-processing mechanism designed for scenarios where a target is obstructed from direct line-of-sight (LOS) observation by occluding structures (e.g., buildings). In such cases, radar transmissions can reach the target via one or multiple specular reflections off building surfaces, and the corresponding reflected echoes can be exploited to infer the presence and kinematic properties of targets within radar shadow regions. This mode therefore enhances target acquisition in visually occluded environments, particularly for objects located behind corners or within building-induced blind zones, as shown in Fig. 5.

The simulation workflow proceeds as follows. First, the scene is instantiated, and the radar’s placement and sensing parameters are configured; after this, the radar waveform is transmitted. Two operating conditions are considered: (i) a clean propagation channel and (ii) a channel with injected environmental interference. The system then checks whether the target lies within the radar’s effective detection range. If the target is out of range, the process terminates with no detection. If the target is in range, the pipeline evaluates whether occluding objects block the direct path and accordingly branches to either direct-path (LOS) detection or reflection-assisted (NLOS) detection along the reflected propagation path. The received echoes are then collected and analyzed, followed by deep-learning-based interference suppression. Finally, waveforms are visualized for inspection, quantitative performance metrics are computed, and detection outcomes from the direct and reflection-based modes are compared.

In this study, the simulation parameters and learning settings were carefully specified for both the radar signal generation environment and the neural-network-based environmental interference mitigation model. The FMCW radar configuration adopts a carrier (center) frequency of 77 GHz and a bandwidth of 200 MHz. The number of sampling points per chirp is set to 300, and the chirp duration is fixed at 5.5 μs to emulate practical automotive radar settings. To control interference severity in a consistent manner, we introduce a global parameter termed the Noise Factor, which is initialized to 0.5 and varied across experiments (e.g., 0.5–3.0) to represent different levels of environmental disturbance. For each simulation run, randomized seed initialization is applied to generate diverse realizations of the synthetic radar signals. This design supports robust evaluation of model stability and generalization across varying noise realizations.

The synthetic waveforms incorporate appropriate propagation time delays and amplitude attenuation to model target reflections. The resulting signals are then corrupted by additive Gaussian noise according to the specified Noise Factor, thereby mimicking signal distortion under environmental interference. For interference mitigation, we adopt a modified AlexNet architecture. The network consists of five convolutional layers followed by three fully connected layers, and it is trained for a regression objective to recover denoised radar signals from inputs contaminated by interference. Using an AlexNet-derived backbone provides a structured baseline that facilitates controlled comparisons between raw and processed waveforms while enabling architecture-level modifications for improved denoising capability. The network is trained using the Adam optimizer, which combines first-order momentum with adaptive second-order learning rates and employs bias-corrected updates. This optimizer is well-suited for noisy learning signals and can improve convergence stability in the presence of interference. The learning rate is set to 0.001, the batch size is 32, and training is performed for 300 epochs. To further improve generalization and reduce ordering effects, the training data are reshuffled after each epoch.

After constructing the simulation scenarios and finalizing the neural network architecture, we evaluate performance using two quantitative metrics: Root Mean Square Error (RMSE) and Signal-to-Noise Ratio (SNR). RMSE is used to measure estimation accuracy, where a lower RMSE indicates reduced prediction error after model enhancement. SNR is used to characterize signal clarity, where higher SNR corresponds to a larger proportion of useful signal energy relative to noise, implying improved interpretability of the radar returns. By jointly analyzing RMSE and SNR under various interference settings (Noise Factor) and network configurations, the effectiveness of the proposed interference mitigation approach can be objectively and consistently validated.

By comparing the baseline AlexNet-based network with the proposed improved model, the differences in handling environmentally induced interference become evident. For the baseline network, introducing interference results in pronounced variations in the error signal, particularly under stronger noise conditions, where the deviation from the clean reference widens considerably. This behavior indicates limited interference tolerance, as denoising provides only partial improvement in waveform clarity, with residual irregular oscillations persisting. Such artifacts suggest that the baseline architecture cannot fully suppress complex interference patterns, which is likely attributable to insufficient representational capacity and limited feature refinement. Consequently, the baseline model is more susceptible to amplitude instability under noisy inputs, which can propagate into downstream errors in range and angular estimation. To illustrate these effects in detail, we select the front radar as a representative example among the three simulated radars, as shown in Fig. 6(1) Because the front radar plays the primary role in forward detection and is typically exposed to the most direct interference in the simulated intersection scenario, it provides a meaningful basis for characterizing signal distortion and evaluating interference mitigation performance. The observed instability in the baseline output highlights the limitations of the original architecture and motivates the need for a more robust model in complex radar environments.

In contrast, the improved neural network exhibits a substantial reduction in error variability across different interference levels. This improvement is mainly attributed to the incorporation of residual connections and complementary optimization strategies. Residual learning promotes more reliable gradient propagation through deeper layers and alleviates degradation effects that can arise in standard deep networks, thereby improving stability when processing interference-contaminated signals. In addition, batch normalization and improved nonlinearities further enhance the model’s ability to attenuate noise and stabilize intermediate feature distributions. As a result, the enhanced model constrains signal fluctuations within a narrower band under various interference settings, demonstrating consistently higher robustness than the baseline architecture. After denoising, the interference-induced error is significantly reduced, and the reconstructed waveform more closely matches the clean reference signal, as shown in Fig. 6(2). Overall, these improvements enable more accurate and stable signal reconstruction under complex interference conditions, providing a stronger foundation for the subsequent estimation of target range and angular parameters. Mechanistically, the residual connection biases the network toward learning a correction term (i.e., the residual error) instead of reconstructing the entire echo from scratch, which improves gradient flow and stabilizes optimization under strong interference. Batch normalization further mitigates internal covariate shift, helping the regressor maintain consistent feature scaling and reconstruction behavior across varying noise levels and clutter conditions.

In this study, radar-based range and angular estimates between the ego vehicle and target objects were obtained from simulation by modeling direct and reflected propagation paths and then computing the corresponding root mean square error (RMSE) with respect to the ground-truth geometry. RMSE serves as the primary quantitative metric, where lower values indicate higher estimation accuracy. To examine the relationship between interference intensity and detection performance, we focus on a severe interference setting (Noise Factor = 3.0). Three conditions are evaluated: (i) a clean environment, (ii) an interference-contaminated environment, and (iii) the denoised output produced by a conventional deep-learning baseline. The resulting RMSE values enable direct comparison of how interference affects radar estimation and how effectively learning-based denoising mitigates this degradation, as summarized in Table 2. Under high interference (Noise Factor = 3.0), the baseline network exhibits substantial performance deterioration, with the largest degradation observed in azimuth estimation (RMSE = 10.77°). This indicates that the baseline model struggles to extract stable angle-related features when the input is heavily corrupted by noise and clutter. The post-denoising results show partial recovery; however, non-negligible residual errors remain under severe interference. Overall, these outcomes highlight key limitations of the conventional architecture, including insufficient robustness to high-intensity noise, limited angle-sensitive feature learning, and reduced adaptability to environmental variations such as occlusion and changes in reflectivity.

By contrast, the improved neural network—augmented with residual learning and batch normalization-achieves consistently lower RMSE across range, azimuth, and deflection angle estimates. The enhanced model demonstrates stronger noise suppression and more stable estimation behavior in challenging conditions, validating the effectiveness of the architectural and training refinements, as reported in Table 3. Under Noise Factor = 3.0, the improved model outperforms the baseline across all radar targets. For the front radar, the baseline model yields a range RMSE of 5.48 m and an azimuth RMSE of 18.95°, reflecting limited tolerance to severe interference. In comparison, the improved model reduces the range RMSE to 0.56 m and the azimuth RMSE to 0.46°, while also lowering the deflection-angle RMSE from 0.93° to 0.73°. These results demonstrate that the proposed architecture maintains stable and accurate estimation under strong interference, providing a more reliable foundation for autonomous vehicle perception and intelligent sensing in occluded urban environments. In this study, radar-based distance and angle estimations between the vehicle and target obstacles were derived through simulation using software tools. These estimations were obtained by modeling signal reflections from surrounding objects, followed by the calculation of the corresponding Root Mean Square Error (RMSE). A lower RMSE value indicates higher estimation accuracy. To examine the relationship between interference intensity and detection accuracy, simulations were conducted under a high interference level (Noise Factor = 3.0). Three scenarios were considered: the original clean environment, the environment with added interference, and the denoised output using a conventional deep learning model. RMSE was employed as the primary evaluation metric to quantitatively compare the detection performance across different conditions. This enabled a clear analysis of the impact of interference on radar signal estimation and the effectiveness of neural network-based denoising approaches, as shown in Table 2. Fig. 7 complements Tables 2 and 3 by providing an at-a-glance visualization of the RMSE reduction and the corresponding spatial convergence toward the ground truth under severe interference.

We further evaluate denoising effectiveness using the signal-to-noise ratio (SNR) measured at three radar units (front, left, and right) under three interference levels (Noise Factor = 0.5, 1.5, and 3.0). For each radar, SNR is compared across two conditions: (i) the denoised output produced by the AlexNet-based baseline model, and (ii) the denoised output produced by the proposed improved network. The results are summarized in Table 4. Overall, the proposed model yields higher SNR than the baseline AlexNet-based model across all radar positions and noise levels, indicating more effective suppression of interference while preserving the underlying signal components. However, the magnitude of SNR improvement decreases as interference intensity increases. Under low interference (Noise Factor = 0.5), denoising yields a pronounced SNR gain, indicating that the dominant corruption can be effectively removed without substantially distorting the target echo structure. In contrast, under severe interference (Noise Factor = 3.0), the SNR gain becomes more modest, implying that high-intensity noise and complex clutter components are more difficult to separate from the signal, even with the improved architecture.

In addition, SNR gains differ across radar locations (front vs. lateral units), suggesting that denoising performance depends on the spatial configuration and the corresponding propagation/interference characteristics. This observation is consistent with intersection scenarios where reflection geometry and occluding structures can produce radar-dependent multipath patterns and non-uniform interference distributions. These findings motivate further investigation into position-aware modeling and adaptive strategies that account for radar placement and environment-specific propagation conditions. The SNR (Signal-to-Noise Ratio) performance of three radar units (front, left, and right) was evaluated under three interference levels (Noise Factor = 0.5, 1.5, 3.0). Comparisons were made between the baseline (no interference), the original Alex Net-based model, and the improved neural network, as shown in Table 4.

To substantiate the deployment-oriented discussion, we report the parameter count, multiply–accumulate operations (MACs), and FP32 weight memory for batch-1 inference on a single 300-sample chirp input. Compared with the baseline AlexNet (1D), the proposed residual + batch-normalization upgrade increases the parameter count from 21.50 to 21.89 M (≈1.81%) and increases MACs from 107.05 to 136.54 M (≈27.54%). Under the common convention that 1 MAC = 2 floating-point operations (FLOPs), this corresponds to 214.10 and 273.08 M FLOPs, respectively. Despite the increased MACs, the FP32 weight memory changes only modestly, as the overall parameter budget remains dominated by the fully connected layers. The computational complexity and memory footprint are summarized in Table 5.

Absolute inference latency is hardware-dependent (CPU/GPU/automotive system-on-chip (SoC)) and also depends on implementation details and the end-to-end radar processing pipeline. In practice, the weight memory footprint can be further reduced using FP16/INT8 quantization; therefore, we frame real-time suitability as platform-specific and leave full embedded latency benchmarking (including preprocessing and postprocessing) as future work.

Although the proposed framework demonstrates the feasibility of reflected-path NLOS detection under controlled simulation settings, several limitations remain before real-world deployment. Future work will focus on (i) unified benchmarking, (ii) improved environmental realism, (iii) deployment efficiency, and (iv) sim-to-real robustness.

First, we will establish a unified benchmarking protocol to systematically compare our method with representative published NLOS radar approaches, including classical signal-processing baselines (e.g., adaptive filtering, wavelet denoising, matrix-pencil methods) and modern learning-based alternatives, using consistent scenarios/datasets and standardized metrics (e.g., RMSE, SNR gain, precision/recall) for fair and reproducible evaluation.