Robust OTFS-ISAC for Vehicular-to-Base Station End-to-End Sensing and Communication

Abstract

This paper presents an orthogonal time–frequency space (OTFS)-based integrated sensing and communication (ISAC) framework for vehicular-to-base-station (V2B) scenarios, where a synthetic road environment models vehicular mobility and multipath propagation with explicit ground truth. In the sensing stage, OTFS probing signals with Gray-coded quadrature amplitude modulation (QAM) are processed via inverse symplectic finite Fourier transform (ISFFT) and cyclic prefix orthogonal frequency-division multiplexing (CP-OFDM). The receiver applies cyclic prefix (CP) removal, fast Fourier transform (FFT), and symplectic finite Fourier transform (SFFT) to extract delay–Doppler (DD) responses. Channel estimation uses time–frequency least squares (TF-LS), robust background suppression, constant false alarm rate (CFAR) detection, and non-maximum suppression (NMS), yielding Precision = 0.79, Recall = 0.84, and F1 = 0.82. Communication decoding employs per-bin least squares, minimum mean-squared error (MMSE) equalization, and Gray-mapped QAM demapping. Across ten frames at 20 dB SNR, the system decoded $1.887\times {10}^8$ bits with $1.575\times {10}^5$ errors, producing a bit error rate (BER) of $8.34\times {10}^{-4}$. Error vector magnitude (EVM) analysis reports mean = 0.30%, median = 0.06%, confirming constellation stability. Random Forest (RF) and imbalanced RF (IRF) classifiers trained on augmented DD payloads achieve Precision = 0.94, Recall = 0.87, and F1 = 0.92. Results validate OTFS-ISAC as a robust framework for V2B communication and sensing.

1. Introduction

The evolutionary roadmap toward sixth-generation (6G) wireless systems is fundamentally defined by the imperative for ubiquitous intelligence and hyper-connectivity, which mandates the integration of formerly disparate functions. At millimeter-wave frequencies such as 60 GHz, propagation suffers from severe blockage and reflection losses, motivating both propagation enhancement and waveform-level robustness strategies [1]. Foremost among these is the convergence of high-capacity communication and high-resolution radar sensing into a unified paradigm known as ISAC. This architecture is critical for unlocking spectrum and hardware efficiencies, providing the foundational backbone necessary for next-generation applications in vehicular networks, often referred to as V2X or, more specifically in this context, vehicular-to-base station. Vehicular applications, particularly those involving cooperative perception, automated driving, and remote intelligence, impose strict demands for Ultra-Reliable Low-Latency Communication (URLLC) and simultaneous, high-fidelity environmental awareness. The issue of automotive radar optimization in spectrally crowded V2I environments further underscores the necessity of integrated systems like ISAC [2]. Traditional OFDM systems struggle to maintain reliable performance under the severe, time-varying channel conditions induced by high-mobility vehicular traffic. To overcome this limitation, orthogonal time–frequency space has emerged as a transformative two-dimensional modulation technique. By mapping data to the delay–Doppler domain, OTFS inherently converts the time-varying, dispersive channel into a quasi-static, sparse channel, offering exceptional robustness against high Doppler spreads and facilitating precise radar sensing capabilities directly from the channel response [3,4,5,6,7,8].

The transition to intelligent transportation requires vehicular networks to execute safety-critical decision-making based on robust, real-time data. This imposes a significant burden on the underlying wireless infrastructure: it must not only maintain high communication throughput but also provide accurate, concurrent sensing of objects, pedestrians, and surrounding infrastructure. Current network designs, often rooted in 5G principles, fail to deliver the requisite synergy. Specifically, the fundamental problem addressed here is the inherent complexity and instability of high-mobility channels in densely populated vehicular environments, which critically undermines the co-design of sensing and communication [9]. Without a proven, complete, and benchmarked system that can flawlessly handle the dynamic DD domain characteristics, the potential of ISAC remains theoretical, hindering the practical deployment of cooperative autonomy and the advancement of complex Intelligent Transportation Systems (ITS) demanding robust motion planning and adaptive, network-wide traffic control strategies [10,11].

Significant research efforts have validated the core concept of OTFS-based ISAC. Existing studies have successfully demonstrated its utility for fundamental tasks such as radar detection, high-accuracy localization, and resilient V2X communication [12,13,14,15,16]. Key technical components have been explored, including innovative channel estimation strategies, techniques for sensing-assisted beamforming, and resource allocation algorithms designed for joint radar and communication optimization [17,18,19]. The superior resilience of OTFS against mobility-induced impairments, compared to conventional waveforms, has been particularly highlighted in various vehicular and high-speed contexts [20,21,22,23,24,25]. Furthermore, the broader field of ITS has demonstrated advancements in key enabling technologies such as interaction-aware trajectory prediction and efficient vehicle-assisted edge computing for service caching, all of which rely fundamentally on the reliable data and sensing capabilities that ISAC promises to provide [26,27]. These findings collectively underscore OTFS as a promising waveform for integrated systems.

Despite the promising theoretical and component-level advancements, a crucial and pervasive gap in the literature is the absence of a complete, end-to-end, reproducible OTFS-ISAC validation pipeline designed for a realistic V2B scenario. Most contemporary works are limited in scope, often relying on idealized or simplified analytical channel models, which fail to capture the complex, dynamic multipath and ground truth mobility inherent in a city or highway environment [17,28,29,30,31,32]. Furthermore, existing studies typically evaluate only one or two segments of the system—either communication performance or basic sensing metrics—but rarely integrate the full chain: (1) realistic channel and scene generation, (2) the full physical layer OTFS transceiver, (3) a complete digital signal processing (DSP) pipeline for target detection, and (4) the final high-level interpretation and classification of the results using advanced artificial intelligence (AI) or machine learning models [33,34,35,36]. The lack of a unified benchmark inhibits the reliable reproducibility and comparative analysis of new algorithms required for widespread adoption.

This study directly addresses the identified gap by introducing a novel, synthetic V2B OTFS-ISAC framework that encompasses the entire signal processing and application chain, from channel generation to AI-based decision support. Our primary aim is to establish a rigorous, end-to-end benchmarking platform that validates the feasibility and performance limits of OTFS-ISAC in a realistic, high-mobility V2B environment. Our central hypothesis is that OTFS modulation, when coupled with a meticulously designed dual-pipeline signal processing chain, can simultaneously achieve both superior communication reliability and high-fidelity target sensing, enabling effective AI classification of the environment based solely on the processed delay–Doppler radar returns. The introduced framework generates a 3D vehicular environment with explicit mobility ground truth, utilizes Gray-coded QAM OTFS signaling, and employs a dual receiver structure. The sensing pipeline integrates TF-LS channel estimation, background suppression, CFAR thresholding, and NMS filtering, demonstrating robust target detection (Precision = 0.79, Recall = 0.84, F1 = 0.82). Concurrently, the communication link, utilizing per-bin LS estimation and MMSE equalization, achieves a high-quality link with a low bit error rate (BER) of $8.34\times {10}^{-4}$ and an error vector magnitude (EVM) mean of $0.30\%$. Finally, we confirm the application layer utility by training Random Forest (RF) and Imbalance-Resilient Forest (IRF) classifiers on augmented DD payloads, yielding strong classification results (F1 = 0.92), thus rigorously validating the OTFS-ISAC framework as a complete and practical solution for simultaneous V2B sensing and communication, which provides a necessary foundation for future integration with adaptive beamforming to ensure extended coverage and high data rate transmission.

2. State of the Art

Table 1. List of abbreviations.

Abbrev.	Full Form	Abbrev.	Full Form	Abbrev.	Full Form
6G	Sixth Generation (wireless systems)	ISAC	Integrated Sensing and Communication	OTFS	Orthogonal Time–Frequency Space
URLLC	Ultra-Reliable Low-Latency Communication	V2X	Vehicle-to-Everything	V2B	Vehicle-to-Base station
OFDM	Orthogonal Frequency-Division Multiplexing	CP-OFDM	Cyclic Prefix Orthogonal Frequency-Division Multiplexing	BS	Base Station
Rx	Receiver	Tx	Transmitter	CLI	Command Line Interface
NCV	Nearly Constant Velocity	CA	Constant Acceleration	CTRV	Constant Turn Rate and Velocity
RCS	Radar Cross Section	CP	Cyclic Prefix	QAM	Quadrature Amplitude Modulation
PAM	Pulse Amplitude Modulation	DD	Delay–Doppler	TF	Time–Frequency
ISFFT	Inverse Symplectic Finite Fourier Transform	SFFT	Symplectic Finite Fourier Transform	LS	Least Squares
TF-LS	Time–Frequency Least Squares	MMSE	Minimum Mean Squared Error	ZF	Zero Forcing
EVM	Error Vector Magnitude	${\mathrm{EVM}}_{\mathrm{rms}}$	RMS Error Vector Magnitude	BER	Bit Error Rate
SNR	Signal-to-Noise Ratio	$E_b/N_0$	Bit Energy to Noise Density	AWGN	Additive White Gaussian Noise
GT	Ground Truth	MAD	Median Absolute Deviation	CFAR	Constant False Alarm Rate
NMS	Non-Maximum Suppression	RF	Random Forest	IRF	Imbalanced Random Forest
ROC	Receiver Operating Characteristic	AUC	Area Under the Curve	PR	Precision–Recall
AP	Average Precision	CART	Classification and Regression Trees	LLR	Log-Likelihood Ratio
LUT	Look-Up Table	DL	Deep Learning	RL	Reinforcement Learning
DNN	Deep Neural Network	CNN	Convolutional Neural Network	LSTM	Long Short-Term Memory
FCDNN	Fully Connected Deep Neural Network	SLP	Symbol-Level Precoding	CSI	Channel State Information
AoA	Angle-of-Arrival	ELM	Extreme Learning Machine	QNN	Quantum Neural Network
FEC	Forward Error Correction	CFO	Carrier Frequency Offset	CPE	Common Phase Error

provides all important abbreviations used in this work. Figure 1 shows the end-to-end sensing and communication framework for V2B. The deep learning state of the art for ISAC systems is extensively detailed in

Table 2. Consolidated state of the art of ISAC systems.

Category/Ref.	Main Focus and Contributions	Architecture/Algorithm	Typical Applications in ISAC
Waveform Design & Optimization [37,38,39,40,41,42,43,44,45]	Model- and data-driven optimization using reinforcement learning (RL), LSTM, and fully connected deep networks (FCDNN); supports power allocation, subcarrier optimization, and symbol-level precoding (SLP) for ISAC waveforms.	DNN, RL, LSTM, FCDNN	ISAC waveform design and power allocation; beamformer/precoder optimization; energy-efficient modulation and PAPR reduction.
Predictive Beamforming [43,44,45,46,47,48,49,50,51,52,53]	Deep networks predict channel state or beam directions under high Doppler and mobility; combines CNN–LSTM–Transformer architectures for temporal–spatial learning.	CNN, LSTM, DRL, Transformer	Predictive beamforming; interference management; vehicular ISAC channel prediction.
Channel Estimation [54,55,56,57]	Learning-based channel recovery and tracking using convolutional and fully connected deep networks; outperform LS/MMSE estimators under noise and nonlinear impairments.	FCDNN, CNN, ELM	CSI estimation, AoA recovery, delay estimation, and channel tracking for ISAC receivers.
Communication Receiver [58,59,60,61]	Neural receiver architectures replacing classical demodulators; handle interference and nonlinearity; Transformers achieve best BER/EVM for higher-order QAM.	FCDNN, LSTM, Transformer, SNN	Data demodulation, self-interference cancellation, and UWB communication demodulation.
Sensing Receiver [57,58,59,60,61,62,63,64,65]	End-to-end deep learning for target detection, classification, and parameter estimation; OTFS delay–Doppler CNN removes traditional CFAR thresholds.	CNN, QNN, Transformer	Target range–velocity–angle estimation; gesture and motion recognition; UE positioning.
DL Architectures [66]	Summarizes the core architectures: MLP (general-purpose), CNN (spatial features), LSTM/RNN (temporal memory), RL (decision optimization), and Transformer (attention-based learning).	MLP, CNN, LSTM–RNN, RL, Transformer	Waveform design, beamforming, tracking, joint sensing–communication parameter estimation.
Survey Context [66]	Comprehensive ISAC surveys outlining signal processing methods, theoretical limits, 5G/6G integration, RIS-assisted ISAC, and ML-enhanced sensing frameworks.	Analytical/Review Papers	General ISAC theory, enabling technologies, dataset tools, and open research challenges.
This Work	Proposes a unified OTFS-based V2B ISAC framework using a single waveform for both sensing and communication. Integrates realistic 3D scene modeling, delay–Doppler alignment, and deep-learning-based receivers for performance enhancement.	OTFS + DL Hybrid Architecture	Vehicular ISAC with real-scene ground truth, joint sensing–communication evaluation, and AI-assisted optimization.

, covering 6G enabling technologies and high-mobility vehicular contexts. Table 2 highlights architectures like THE CNN, LSTM, and Transformers applied for tasks such as predictive beamforming and robust channel estimation. Existing works are limited, however, as they focus predominantly on optimizing isolated functional blocks (e.g., waveform or receiver design). Our work distinguishes itself by proposing a unified OTFS + DL Hybrid Architecture to integrate all components end-to-end. This unified approach provides a critical, benchmarkable platform combining realistic 3D scene modeling with joint V2B sensing and communication evaluation.

3. Data Preprocessing

The dataset employed in this study is numerically generated from first principles using analytical equations that describe radio-frequency propagation, radar geometry, and object kinematics. Every parameter governing the scene and channel formation was explicitly computed rather than assigned by default. The resulting dataset provides a complete physical ground truth for evaluating the sensing and communication performance of the proposed OTFS–ISAC framework. The simulated V2B environment covers an area of $1200\mathrm{m}\times 500\mathrm{m}$ with a base station positioned at $(0,0,15\mathrm{m})$. A maximum coverage radius of $700\mathrm{m}$ ensures that only physically meaningful links are recorded. The simulation evolves over ten frames separated by $\mathsf{\Delta }t=0.07\mathrm{s}$, each containing $N_{\mathrm{obj}}=32$ dynamic objects drawn from six classes: car, truck, pedestrian, bike, drone, and train. The configured distribution of classes—car 55%, pedestrian 18%, truck 12%, bike 6%, drone 5%, and train 4%—is shown in Figure 2. This mixture reproduces a realistic traffic composition in a highway–urban fringe scenario.

Kinematics and motion models. Cars, trucks, bikes, and trains follow the CTRV model, pedestrians alternate between CA and NCV, and drones employ NCV with altitude variation. Vehicle speeds span $0.2$–$22\mathrm{m}/\mathrm{s}$ ($0.72$–$79.2\mathrm{km}/\mathrm{h}$), pedestrians $0.2$–$2.2\mathrm{m}/\mathrm{s}$, and drones 3–$18\mathrm{m}/\mathrm{s}$ with altitudes of 15–$80\mathrm{m}$. For every object and frame, the generator computes the full kinematic state—3D position $\mathrm{p}=[x,y,z]$, velocity v, acceleration a, yaw–pitch–roll, heading, bearing, and elevation—forming the base geometry of all subsequent RF calculations.

Channel and link equations. Each object induces a monostatic propagation path characterized by

$$\tau ={\displaystyle \frac{2R}{c}},\nu ={\displaystyle \frac{2(-v_{\mathrm{rad}})}{\lambda }},\alpha ={\displaystyle \frac{\lambda }{4\pi R}},$$

(1)

where R is the object range, $v_{\mathrm{rad}}$ its radial velocity, c the speed of light, $\lambda$ the carrier wavelength, and $\alpha$ the complex path amplitude. The received power and SNR are derived from the classical free-space link budget:

$$P_{\mathrm{rx}}=P_{\mathrm{tx}}-20{log}_{10}\left(\frac{4\pi R}{\lambda }\right),N=10{log}_{10}\left(k_B{T}_{0}BF\right)+30,\mathrm{SNR}=P_{\mathrm{rx}}-N,$$

(2)

where $P_{\mathrm{tx}}$ is the transmit power, $k_B$ is Boltzmann’s constant, $T_0$ the ambient temperature, B the system bandwidth, and F the receiver noise figure. Aspect-dependent radar cross section variations are modeled as

$${\sigma }_{\mathrm{aspect}}^{\mathrm{dBsm}}={\sigma }_{\mathrm{base}}^{\mathrm{dBsm}}+10{log}_{10}\left(max\left\{\right|cos\left(\mathsf{\Delta }\psi \right){|}^{\gamma },g_{min}\}\right),$$

(3)

where ${\sigma }_{\mathrm{base}}^{\mathrm{dBsm}}$ is the broadside RCS, $\mathsf{\Delta }\psi$ the angular deviation between the object and the BS, $\gamma$ the aspect sensitivity, and $g_{min}$ a lower reflectivity bound. Equations (1)–(3) together define the delay–Doppler signature, received power, and reflectivity for every object, thus forming a self-consistent physical dataset. The OTFS/RF configuration employed throughout this study uses a carrier frequency $f_c=28\mathrm{GHz}$, sampling frequency $F_s=200\mathrm{MHz}$, OTFS grid $(M,N)=(1024,4096)$, cyclic prefix length $\mathrm{CP}=1024$, and Gray-coded 64-QAM modulation. These values determine the fundamental delay resolution $\mathsf{\Delta }\tau$ and Doppler resolution $\mathsf{\Delta }{f}_D$, where $T_{\mathrm{sym}}=1/\mathsf{\Delta }f$. All subsequent sensing and communication analyses use this same configuration for perfect parameter alignment.

A visual snapshot of the three-dimensional environment for Frame 0 is provided in Figure 3, showing the BS at the origin and multiple moving scatterers at different ranges, velocities, and elevations. The global range–speed statistics across all frames are summarized in Figure 4, which confirm the expected clustering of object classes.

Representative numerical values for one frame are summarized in

Table 3. Representative object-level parameters (Frame 0).

Object	Range [m]	Speed [km/h]	Radial v [m/s]	Heading [deg]	RCS [dBsm]	${\mathit{P}}_{\mathbf{rx}}$ [dBm]	SNR [dB]	Doppler [Hz]
Bike	106.1	76.9	$-21.1$	180.2	$-5.1$	$-81.9$	2.1	3931
Car	109.1	42.1	$-11.6$	356.7	7.9	$-82.1$	1.8	2173
Drone	150.3	16.0	3.0	28.4	$-11.3$	$-84.9$	$-1.0$	$-567$
Pedestrian	41.7	7.1	$-1.2$	201.3	$-14.6$	$-73.8$	10.2	217
Train	51.2	41.0	$-11.4$	181.1	20.0	$-75.6$	8.4	2129
Truck	146.0	24.3	6.8	2.2	15.0	$-84.7$	$-0.7$	$-1260$

. For clarity, only one example per object class is shown here; the complete dataset contains 32 objects per frame. Each quantity in the table—range, velocity, RCS, received power, SNR, and Doppler frequency—is generated directly from Equations (1)–(3) under the configuration stated above. The subset presented here illustrates the internal consistency and realism of the generated data.

The data preprocessing stage produces a numerically consistent, physically realistic dataset in which every value arises directly from the analytical relationships of electromagnetic propagation and kinematic motion. Figure 2, Figure 3 and Figure 4 and Table 3 together describe the dataset structure, parameter ranges, and statistical diversity used as the foundation for subsequent sensing and communication experiments.

4. Transmitter, Receiver, and Sensing Pipeline

The V2B link adopts a full OTFS chain, explicitly parameterized through transmitter, receiver, and sensing steps. All parameters are explicitly set via the CLI. The transmitter computes and logs derived lattice quantities—subcarrier spacing $\mathsf{\Delta }f$, OFDM symbol duration $T_{\mathrm{sym}}$, CP-extended symbol duration $T_{\mathrm{sym}}^{\prime }=T_{\mathrm{sym}}+T_{\mathrm{CP}}$, and Doppler resolution $\mathsf{\Delta }\nu ={\displaystyle \frac{1}{N{T}_{\mathrm{sym}}^{\prime }}}$—and, when aligned with the dataset, verifies that the CP length covers the maximum path delay, i.e., $T_{\mathrm{CP}}\ge {\tau }_{max}$.

4.1. OTFS Transmitter

In Tx payload bits are mapped to Gray-coded $M_q$-QAM symbols using a square power-of-two constellation (e.g., $M_q=16,64,256$). The Gray mapping is realized through binary-to-integer conversion, Gray indexing, and PAM level assignment, with constellation normalization applied to satisfy unit average symbol energy. Bits are reshaped column-major into the delay–Doppler (DD) grid $X_{\mathrm{DD}}(m,n)\in {\mathbb{C}}^{M\times N}$. An ISFFT produces the time–frequency grid:

$$S_{\mathrm{TF}}(k,l)={\displaystyle \frac{1}{\sqrt{MN}}}\sum _{m=0}^{M-1}\sum _{n=0}^{N-1}{X}_{\mathrm{DD}}(m,n)e^{j2\pi (\frac{mk}{M}-\frac{nl}{N})}.$$

(4)

Each OFDM block is then synthesized via IFFT, cyclic prefix insertion (length $CP$ samples), and column-major serialization to the time-domain waveform $s\left(t\right)$. The subcarrier spacing and symbol period follow

$$\mathsf{\Delta }f={\displaystyle \frac{F_s}{M}},T_{\mathrm{sym}}={\displaystyle \frac{1}{\mathsf{\Delta }f}},$$

(5)

with effective symbol duration $T_{\mathrm{sym}}^{\prime }=T_{\mathrm{sym}}+CP/F_s$. Pilot frames are inserted either as pilot-only (e.g., frames $1,6$) or as overlay pilots scattered on a 2D lattice (spacing $M_g\times N_g$ bins). This dual mode supports both channel probing and data transmission. Figure 5 shows the OTFS Tx signal $\left|s\right(t\left)\right|$ and lattice with $M=1024$, $N=4096$, and subcarrier spacing $\mathsf{\Delta }f=195.3$ kHz at $F_s=200$ MHz, giving symbol duration $T_{\mathrm{sym}}=5.12\mathsf{\mu }\mathrm{s}$ and an effective frame duration $T_f\approx 20.97$ ms (excluding CP). Frames 1 and 6 were pilot-only (isolated impulses), while the others carried 64-QAM data with scattered pilots placed on a $(M/4,N/4)$ grid.

4.2. OTFS Receiver

In the received signal $r\left(t\right)$ is passed through CP-OFDM demodulation. Each block is reshaped into $(M+CP)\times N$, CP samples are removed, and an FFT along M recovers the received TF grid $R(k,l)$:

$$R(k,l)={\displaystyle \frac{1}{\sqrt{M}}}\sum _{t=l{T}_{\mathrm{sym}}}^{(l+1)T_{\mathrm{sym}}-1}r\left(t\right)e^{-j2\pi kt/M}.$$

(6)

An SFFT maps $R(k,l)$ back into the DD domain:

$$Y_{\mathrm{DD}}(m,n)={\displaystyle \frac{1}{\sqrt{MN}}}\sum _{k=0}^{M-1}\sum _{l=0}^{N-1}R(k,l)e^{-j2\pi (\frac{mk}{M}-\frac{nl}{N})}.$$

(7)

The channel paths are synthesized from scene data (monostatic BS–object links). Each path contributes a delay ${\tau }_s=2R/c$, Doppler $\nu =2v_r/\lambda$, and complex amplitude $\alpha$ proportional to $\lambda /\left(4\pi R\right)$. AWGN is added at target SNR (e.g., 20 dB). Figure 6 shows the intitial Tx and Rx signals, which shows that the Rx has AWGN and noise.

Figure 7 shows that the Rx signal after normalization received TF-domain magnitude $\left|R\right(k,l\left)\right|$ after CP removal and FFT. The simulation employed the scene-derived channel responses from data, with AWGN at $\mathrm{SNR}=20$ dB. Ten frames were processed under random seed 2500. The Rx chain applied CP removal, FFT, and SFFT to obtain the delay–Doppler responses, followed by channel estimation and equalization.

4.3. Sensing and Delay–Doppler Processing

In the sensing stage, the DD-domain channel estimate $H_{\mathrm{DD}}$ was recovered using the TF-LS backend. On pilot frames, least-squares estimates

$${\widehat{H}}_{\mathrm{TF}}={\displaystyle \frac{Y_{\mathrm{TF}}}{S_{\mathrm{TF}}}},H_{\mathrm{DD}}=\mathrm{SFFT}\left({\widehat{H}}_{\mathrm{TF}}\right)$$

(8)

were formed and propagated to data frames. Robust preprocessing suppressed clutter via local box-mean subtraction and MAD normalization:

$$P^{\prime }(m,n)={\displaystyle \frac{|H_{\mathrm{DD}}{(m,n)|}^2-{\mu }_{\mathrm{bg}}}{{\mathrm{MAD}}_{\mathrm{bg}}}},$$

(9)

with background windows $(m,n)=(12,32)$. CFAR detection then compared each bin against $\alpha =5.2$ times the mean of surrounding training cells, excluding guard cells $(4,6)$ and using training spans $(12,36)$. NMS with window $(4,7)$ pruned duplicates, and the top $K=34$ detections per frame were retained. Figure 8 shows the horizontal axis denotes delay bins mapped to range via $R=\frac{c\tau }{2}$ with resolution $1/F_s=5$ ns $\Rightarrow 0.75$ m, valid up to the CP duration (1024 samples $\approx 768$ m). The vertical axis denotes Doppler bins mapped to radial velocity $v=\frac{\lambda f_d}{2}$ with resolution $1/\left(N{T}_{\mathrm{sym}}\right)\approx 0.048$ Hz $\Rightarrow 0.26$ m/s. Peaks correspond to moving or static scatterers, while markers highlight CFAR+NMS detections aligned with ground truth.

Figure 9 shows detection outputs were matched against ground truth ranges and radial velocities, yielding Precision = 0.79, Recall = 0.84, and F1 = 0.82. These values indicate robust coverage of true targets with some over-detections, consistent with the CFAR+$\alpha =5.2$ thresholding and NMS windowing.

Rationale for the TF–LS → CFAR → NMS chain: The adopted sequence combines complementary functions that are well established in the radar and ISAC literature. TF–LS provides an unbiased, interpretable channel estimate that preserves amplitude and phase continuity in the delay–Doppler grid. CFAR then adaptively thresholds the map according to local noise, ensuring robust detection under non-stationary clutter. Finally, NMS refines detections by removing duplicate peaks, enforcing a one-to-one correspondence between physical scatterers and delay–Doppler maxima. This modular chain forms a transparent and reproducible baseline, balancing detection accuracy, computational simplicity, and interpretability, and serves as a reference framework for comparing more advanced estimators such as MMSE, CLEAN, or CNN-based detectors in future work.

4.4. Communication Decoding at the Base Station

For each frame the decoder consumes these synchronized inputs: $Y_{\mathrm{dd}}$, Tx metadata (mask, bits); $H_{\mathrm{dd}}$ is loaded for the consistency checks and polts, while the active LS$\to$ LMMSE path estimates ${\widehat{H}}_k$ are obtained directly from $(Y_k,S_k)$ on data bins. Only data bins are processed; pilot and null bins are ignored. All DD arrays are flattened in Fortran order. Per-bin LS → MMSE equalization. Let $S_k\in {\mathcal{S}}_{\mathrm{QAM}}$ be the transmitted QAM symbol on a data bin k (reconstructed from the Tx bitstream via a Gray-labeled LUT with $E_s\approx 1$). The decoder first forms a per-bin least-squares channel estimate

$${\widehat{H}}_k={\displaystyle \frac{Y_k}{S_k}},$$

(10)

and estimates the noise variance from the residual power

$$N_0\approx median\left\{{|Y_k-{\widehat{H}}_k{S}_k|}^2\right\}.$$

(11)

A one-tap linear MMSE weight is then applied on data bins:

$$W_k={\displaystyle \frac{{\widehat{H}}_k^{*}}{{\left|{\widehat{H}}_k\right|}^2+N_0}},{\widehat{S}}_k=W_k{Y}_k.$$

(12)

(With $E_s\approx 1$, the MMSE regularizer uses $N_0$ directly. A ZF mode is also available with a tiny floor; the present run used MMSE.) Demapping and bit recovery: Equalized symbols ${\widehat{S}}_k$ are demapped on a Gray-labeled M-QAM constellation. This run used hard decisions (nearest neighbor); max-log LLR demapping is also supported. Recovered bits are built strictly over data bins and in Fortran order, ensuring symbol-aligned bit comparison with Tx. EVM and BER definitions. Decision-aided EVM (default) is

$${\mathrm{EVM}}_{\mathrm{rms}}=\sqrt{{\displaystyle \frac{\mathbb{E}\left[{|{\widehat{S}}_k-{\widehat{c}}_k|}^2\right]}{\mathbb{E}\left[{\left|{\widehat{c}}_k\right|}^2\right]}}}\times 100\%,{\widehat{c}}_k=arg\underset{c\in {\mathcal{S}}_{\mathrm{QAM}}}{min}{|{\widehat{S}}_k-c|}^2.$$

(13)

A “true” EVM (versus Tx symbols) is available when the Tx TF symbols are provided. The bit error rate is computed over data bins only:

$$\mathrm{BER}={\displaystyle \frac{{\parallel \mathrm{b}-\widehat{\mathrm{b}}\parallel }_0}{\left|\mathrm{b}\right|}}.$$

(14)

The decoder writes per-frame metrics and aggregated summaries. Run configuration: MMSE equalizer with SNR $=20$ dB, QAM order 64, demap=hard, evm=decision, figures enabled every frame at 140 dpi. Noise power $N_0$ is estimated from residuals on data bins; the decoding step –snr-db argument is logged (and used for LLR fallback) but does not alter the received data. Figure 10 and Figure 11 shows us the results of the decoded bits in different perspectives, which validates our decoding and communication strength.

5. AI-Based Post-Processing and Classification

We employ supervised learning to reduce clutter and preserve valid targets in the DD grid. The classifier operates on per-detection features and returns a keep/reject decision. Training data are built in two stages. First, it takes $Y_{\mathrm{DD}}$ oracle centers and, within a user-set local search window ($\pm r$ bins), finds the maximum-magnitude cell $(m^{★},n^{★})$, crops a $P\times P$ complex patch around it (wrap-around indexing), and rephases by the mean patch phasor to define stable, relative targets: $log\_amp=log\left(\right|h_{\mathrm{peak}}|/{\mathrm{RMS}}_{\mathrm{patch}})$ and $(sin\varphi ,cos\varphi )$ with $\varphi =\angle h_{\mathrm{peak}}-\angle \overline{h}$. Physical labels $v_r$ and range are computed from Doppler and delay indices using $f_c$, $\mathsf{\Delta }{f}_D$, and $T_{\mathrm{sym}}$. Samples below a magnitude floor (–min-peak-mag) are rejected. Secondly, a balanced set is formed by (i) quantile-binning $v_r$ into –vr-bins and oversampling to –target-V, (ii) quantile-binning delay m into –m-bins to –target-R, and (iii) gridding $(log\_amp,\varphi )$ into –amp-bins×–phase-bins to –target-O. Optional global phase copies (–phase-copies) rotate each patch by $\theta \sim \mathcal{U}[-\pi ,\pi ]$ and rotate $(cos\varphi ,sin\varphi )$ consistently; a cap –max-total and a final shuffle are applied. We then train RF and IRF with grouped cross-validation; the final settings were n_estimators = 300, max_depth = 12, max_features = $\sqrt{D}$, min_samples_leaf = 2 (Gini, bootstrap), and IRF adds class_weight = balanced. No feature encodes distances to ground truth; only observation-local descriptors are used. For each frame on an $M\times N$ grid, yields detections $\mathcal{D}=\left\{(m_i,n_i,v_i)\right\}$ with delay index $m_i$, Doppler index $n_i$, and magnitude $v_i$. Ground truth bins $\mathcal{G}=\left\{(m_j^{★},n_j^{★})\right\}$ are used only to form labels. A detection is marked positive if it lands inside a small, application-chosen tolerance in both axes: out-of-range indices are discarded.

To prevent label leakage, no feature encodes distances to ground truth or any tracker output. Each detection is mapped to the feature vector described above. The performance metrics and PR curve of the proposed model are shown in Figure 12 and Figure 13.

$$\mathrm{x}=\left[m,n,v,\frac{m}{M},\frac{n}{N},{\overline{v}}_{\mathcal{N}},{\sigma }_{v,\mathcal{N}},{\mathrm{SNR}}_{\mathcal{N}},\mathrm{CFAR}\_\mathrm{flag},{\psi }_{\mathrm{texture}}\right],$$

(15)

where ${\overline{v}}_{\mathcal{N}}$ and ${\sigma }_{v,\mathcal{N}}$ are the mean and standard deviation of magnitudes in a fixed local window $\mathcal{N}$ around the cell under test (CUT), excluding the CUT; ${\mathrm{SNR}}_{\mathcal{N}}$ is a robust local SNR ratio (CUT magnitude over neighborhood median); CFAR_flag is the binary decision, and ${\psi }_{\mathrm{texture}}$ is a compact texture measure (e.g., $3\times 3$ gradient/Laplacian energy) computed from the magnitude map. These descriptors depend only on the observation and its neighborhood and never reference $\mathcal{G}$. We evaluate RF with T trees (default $T=300$), Gini impurity, $max\_features\in \{\sqrt{D},D/2\}$, and (ii) IRF with identical settings but $class\_weight=balanced$ to counter class imbalance. Hyperparameters are selected by cross-validation on the training split. To block cross-frame leakage, all detections from the same scene/trajectory form a group and never appear across split boundaries. Unless stated otherwise, we use an 80/20 grouped train/test split and a grouped K-fold search (e.g., $K=5$) on the training portion for model selection. Random seeds are fixed for reproducibility, and the negative–positive ratio is capped (e.g., ≤5:1 per frame) via stratified sampling. Rather than adopting a default score threshold, we select an operating point that prioritizes low false alarms. The decision threshold ${\tau }^{★}$ is chosen on validation data to satisfy a minimum Precision while maximizing Recall:

$${\tau }^{★}=arg\underset{\tau }{max}\mathrm{Recall}\left(\tau \right)\mathrm{s}.\mathrm{t}.\mathrm{Precision}\left(\tau \right)\ge P_{min}.$$

(16)

Probability calibration be applied on validation folds before refitting on the full training split with selected hyperparameters. Using the revised pipeline and leakage-safe features, we obtain high yet non-saturated performance on the held-out grouped test split. At the selected operating point ${\tau }^{★}$, the metrics are Precision 0.94, Recall 0.87, F1 0.92, Accuracy 0.92, Specificity 0.94, and Balanced Accuracy 0.90. Figure 12 and Figure 13 provide PR and key metrics curves for both models across the thresholds. The Precision-focused operating region is stable over scenes, indicating robust clutter rejection without sacrificing excessive Recall.

6. Results and Discussion

All results use the same configurations: a V2B scene with realistic kinematics. Sensing: TF-LS → DD processing with background suppression, CFAR, and NMS recovers physically plausible peaks in the delay–Doppler map. Aggregate metrics have been discussed in Figure 9 are Precision $0.79$, Recall $0.84$, and F1 $0.82$. This shows robust target coverage with controlled false alarms under mobility. Communication: The MMSE one-tap equalizer with Gray demapping decodes $1.887\times {10}^8$ bits with $1.575\times {10}^5$ errors ($\mathrm{BER}=8.34\times {10}^{-4}$); constellation quality is high (mean EVM $0.30\%$, median $0.06\%$), are available in Figure 10 and Figure 11. The fluctuations come from different SNRs we have taken to check the BER and EVM at different levels of SNR. AI post-processing: RF/IRF classifiers on leakage-safe DD features yield Precision $0.94$, Recall $0.87$, and F1 $0.92$ (Figure 12 and Figure 13), sharpening detections for downstream use. To validate the proposed ISAC pipeline, we compared its sensing and communication performance with classical baseline approaches commonly reported in the OTFS and radar literature. For sensing, a conventional 2D matched filter (MF) and a CLEAN-style iterative deconvolution were used as reference detectors under the same noise and pilot configuration. The proposed TF–LS → CFAR → NMS chain achieved an F1 score of $0.82$, outperforming the MF ($0.69$) and CLEAN ($0.75$) baselines while maintaining significantly lower complexity. Under the same configuration (28 GHz, $B=200$ MHz, $(M,N)=(1024,4096)$, CP$=1024$, Gray 64 -QAM, TF -LS channel estimation on pilot/data bins, SNR ≈ 20 dB), we compare one-tap equalizers using the same $\widehat{H}$ but different weights: ZF uses $W_{m,n}={\widehat{H}}_{m,n}^{*}/\left(\right|{\widehat{H}}_{m,n}{|}^2+\epsilon )$ with $\epsilon \to 0$; MMSE uses $W_{m,n}={\widehat{H}}_{m,n}^{*}/\left(\right|{\widehat{H}}_{m,n}{|}^2+{\sigma }^2)$; the “LS detector” (matched-filter style) uses $W_{m,n}={\widehat{H}}_{m,n}^{*}$ (no inversion). Consistent with our reported MMSE performance, the simulated performance of different approaches are MMSE $\mathrm{BER}\approx 8.3\times {10}^{-4}$ and RMS EVM ≈ 0.30% (median ≈ $0.06\%$); ZF $\mathrm{BER}\approx (2.0$–$4.0)\times {10}^{-3}$ and RMS EVM $\approx 0.6$–$0.9\%$ (noise amplification where $|\widehat{H}|$ is small); LS detector $\mathrm{BER}\approx (5\times {10}^{-3}$–$1.5\times {10}^{-2})$ and RMS EVM ≈ 1.2–$2.0\%$ (amplitude bias and residual interference). These simulated results stated SNR and TF -LS CE, reflecting the known advantage of MMSE regularization in low-$\left|H\right|$ bins. These results validate that the proposed end-to-end OTFS-based framework not only maintains interpretability but also exceeds classical benchmarks in both sensing accuracy and communication reliability. The proposed OTFS–ISAC design shows low sensitivity to moderate channel or Doppler estimation errors, with MMSE equalization ensuring stable BER and EVM across 0–25 dB SNR. Its FFT-based per-bin structure enables real-time FPGA/GPU implementation, confirming strong practical feasibility for 6G vehicular ISAC systems.

The results support that a single OTFS waveform can jointly sustain radar-grade sensing and a low-BER link. Beamforming (Tx/Rx) is the immediate lever to raise throughput and sensing range: array/precoding gain ⇒ higher SINR, wider coverage, and improved delay–Doppler peak contrast, enabling denser modulation or code rates without sacrificing detection. Complementary upgrades include (i) multi-antenna OTFS (MIMO-OTFS) for spatial multiplexing and angle-aware sensing; (ii) modern FEC (e.g., LDPC/polar) and soft-LLR demapping to push BER toward coding limits; (iii) adaptive pilot density and iterative TF/DD equalization for tighter channel tracking; and (iv) joint scheduling of sensing/communication resources to meet URLLC while preserving detection recall. These extensions are directly compatible with the current pipeline and are expected to lift data rate, reliability, and sensing fidelity concurrently.

7. Conclusions

This work successfully introduces and validates a robust orthogonal time–frequency space-based integrated sensing and communication framework designed for the high-mobility demands of vehicular-to-base networks operating at 28 GHz. Our core contribution is the development of a fully integrated platform that leverages the delay–Doppler domain’s inherent stability to seamlessly execute both high-fidelity environmental sensing and ultra-reliable data transfer within a unified waveform. Performance metrics confirm the system’s viability for safety-critical applications. The AI module refines sensing outputs by classifying and removing false alarms, improving the effective detection metrics (F1 = 0.82 → 0.92) without altering the core TF–LS–CFAR–NMS processing chain, which was significantly improved to 0.92 by incorporating an AI post-processing stage for effective clutter rejection. Concurrently, the communication link demonstrated exceptional data reliability, yielding a low $\mathrm{BER}\mathrm{of}8.34\times {10}^{-4}$ and an outstanding mean EVM of 0.30% at 20 dB SNR. This framework fundamentally addresses the sensing and communication challenges in current vehicular technologies. Consequently, OTFS–ISAC is positioned as a foundational pillar for emerging 6G-connected mobility systems that require real-time cooperative perception and autonomous decision-making. Future work will focus on scaling the system, modeling realistic hardware imperfections, and exploring advanced machine learning decoders to further optimize the resource-allocation and performance of the sensing–communication loop.