ABSTRACT:

Detecting multiple moving targets and determining their positions and velocities is a
challenging task for radar systems, especially under weak or noisy signals. This work focuses on
a continuous-wave monostatic OFDM radar, which transmits continuous OFDM signals and
measures reflections from targets. A residual convolutional neural network (CNN) is used to
directly estimate the range and velocity of multiple targets from radar range-Doppler maps. The
network treats the task as an image-to-image problem, transforming noisy radar signals into
clear ground-truth maps showing target positions. Synthetic data includes multiple targets with
varying random distances, speeds, and reflection strengths, combined with noise to simulate
real radar conditions. Residual blocks enable training deeper networks without losing accuracy
and enhance precise target detection. The model is tested across different signal-to-noise ratios
(SNRs) and evaluated using root mean square error (RMSE) for range and velocity, and peak
signal-to-noise ratio (PSNR) for map quality. Results demonstrate accurate target localization
and effective noise reduction. Moreover, compared to traditional radar methods and other
deep learning models, this approach achieves higher accuracy, efficiently handles noisy signals,
and supports real-time multi target detection.

INTRODUCTION:
Radar systems are used to detect and track objects by sending electromagnetic waves and
analyzing the reflected signals. They can measure the distance (range) and speed (velocity) of
targets. Among radar types, continuous-wave (CW) radar is a type of radar system where a
known stable frequency continuous wave radio energy is transmitted and then received from any
reflecting objects. Individual objects can be detected using the Doppler effect, which causes the
received signal to have a different frequency from the transmitted signal. When CW radar is
combined with orthogonal frequency-division multiplexing (OFDM), it can also measure target
distances with high resolution This makes continuous-wave Monostatic OFDM radar useful for
applications requiring the detection of multiple moving targets in complex environments, such as
automotive radar for autonomous vehicles, UAV tracking, and dense network scenarios with
drones and other connected devices. Detecting multiple targets in such environments is
challenging due to weak reflections, overlapping signals, and noise that reduces detection
accuracy. Traditional radar methods, such as the 2D periodogram , Multiple Signal Classification
(MUSIC), and Estimation of Signal Parameters via Rotational Invariance Techniques(ESPRIT),
require large amounts of data, have high computational costs, and often perform poorly in low
signal-to-noise ratio (SNR) conditions. They may also be too slow for real-time multitarget
detection. The study addresses these challenges by applying a deep learning approach using a
residual convolutional neural network (CNN) to process radar range-Doppler maps. The CNN
directly estimates the range and velocity of multiple targets while handling noisy signals and
overlapping [Link] using this method detection accuracy can be improved, noise can be
reduced, and faster processing can be provided as compared to other traditional methods.
LITERATURE REVIEW:
Deep learning (DL) has recently emerged as an effective tool for radar signal processing,
particularly for target detection, parameter estimation, and interference mitigation. Traditional
methods like periodograms, Multiple Signal Classification (MUSIC), and Estimation of Signal
Parameters via Rotational Invariance Techniques (ESPRIT) provide accurate parameter
estimates under ideal conditions but often struggle with noise, low SNR, and multiple closely
spaced targets [1]. For interference mitigation, Liu et al. proposed Radar-STDA, a spatial-
temporal denoising autoencoder for FMCW radars that reduces interference while maintaining
high throughput [2]. Similarly, Brown et al. applied Convolutional Neural Network(CNN) based
autoencoders to radar signals, effectively suppressing narrowband and wideband interference
while preserving range accuracy [2].In target detection and recognition, Han et al. used CNNs on
FMCW range-Doppler images to detect small drones, improving detection for weak signals [3].
Dai et al. transformed 1D radar time-series into 2D representations using CNNs, enabling
detection under low SNR conditions [4]

Deep learning can also help classify radar signals and separate overlapping signals. For example,
Kocamış et al. used DL to identify different radar signals even when they overlap, which is
useful in electronic warfare [5]. Similarly, Yavuz developed RadCNN to improve radar detection
in pulsed Doppler radar, especially when the signal is weak or noisy [5].

Delamou et al. [1] introduced a new deep learning approach to improve radar detection when
there are multiple targets. The model uses convolutional neural networks (CNNs) to find both the
distance and speed of multiple targets at the same time. It works directly with radar range-
Doppler maps, which are images showing where targets are and how fast they are moving. Older
methods such as periodograms, Multiple Signal Classification (MUSIC), and Estimation of
Signal Parameters via Rotational Invariance Techniques (ESPRIT) work well in ideal conditions
but struggle when there is involvement of noise or when targets are very close [Link]
CNN-based model provides better accuracy, clearer results, and faster predictions than older
deep learning models like VGG-19 and 2DResFreq.

The network uses residual connections, that helps prevent accuracy from dropping when the
network has many layers. This ensures the model learns effectively even as it becomes deeper.
Unlike older deep learning models, which usually detect only one target at a time, this model can
handle complex environments with many targets. It does not require special antennas and works
in realistic radar scenarios with closely spaced moving objects.

Most previous studies focused on FMCW or pulsed radars, but this CNN works directly with
OFDM radar for range-Doppler estimation. It can estimate multiple targets at once, stays
accurate even under noisy conditions, and makes predictions quickly. Overall, this approach
makes radar systems more reliable, especially in challenging conditions like low signal-to-noise
ratio or multiple moving targets. It shows how deep learning can exceed traditional radar
algorithms in complex environments.
METHADOLOGY:
The selected radar is Continuous Wave (CW) OFDM Radar, The radar continuously transmits
orthogonal frequency subcarriers instead of pulses. Each subcarrier reflects from targets, giving
range and velocity information simultaneously. The received signals are processed into range-
Doppler maps, which show the distance and speed of targets. OFDM allows high resolution
detection of multiple targets, even in noisy environments.

SYSTEM ILLUSTRATION:
This figure illustrates the physical setup and signal interaction of a Monostatic Continuous Wave
(CW) Orthogonal Frequency-Division Multiplexing (OFDM) radar system.

1. Monostatic Radar: The system uses a single antenna (or co-located antennas) for both
transmitting and receiving.
2. Signal Transmission: The radar transmits a high-bandwidth, OFDM signal into the
environment. OFDM is chosen for its superior range resolution and robustness against
multipath effects.
3. Target Interaction: The transmitted signal encounters a Moving Target. Due to the target's
motion and distance, the reflected signal experiences both a time delay (indicating range) and
a Doppler frequency shift (indicating velocity).
4. Signal Reception: The Reflected signal magnitude and phase—which encode the target's
range and velocity information is captured by the same Monostatic CW OFDM radar

Figure #1:Monostatic OFDM Radar System For Target Sensing

1. Input Data Acquisition: The process begins with the Received Signal captured by the
Monostatic Continuous Wave OFDM radar.
2. Range Doppler Map Generation: The received complex signal is first processed (typically
using 2D FFT) to create the Range Doppler Map (RDM). This map is a fundamental radar
output where the target's energy is located at specific coordinates corresponding to its range
(distance) and Doppler frequency (velocity).The RDM data is formatted as a 2-channel input
(I/Q data, or magnitude/phase) along with the Ground Truth (GT) map for training purposes.
3. Residual CNN Processing: The RDM is fed into a Residual Convolutional Neural Network (CNN).
This deep architecture is designed to learn complex, non-linear mappings between the raw RDM data
and the target [Link] network employs three Residual Blocks to extract features, expanding the
channel count from 2 to 32, and finally to 64. Residual connections help stabilize the training of deep
layers. The final Conv2D (64 out 1) layer compresses the rich features back into a single-channel
output [Link] final output is the Predicted CNN map Estimating Range and Velocity Indices. This
map highlights the locations of the targets, effectively performing target detection and parameter
estimation simultaneously.

Fig#2: Deep Learning Based Range and Velocity Estimation using Residual CNN