⚡ electrical engineering

Adaptive High-Speed Radar Signal Processing Architecture for 3D Localization of Multiple Targets on System on Chip

This paper presents an adaptive, hardware-software co-designed 3D radar signal processing accelerator on a Zynq MPSoC that dynamically switches between high-accuracy and low-complexity frameworks based on signal conditions, achieving a 5.6x speedup and a 24% improvement in 6G ISAC communication throughput without increasing hardware complexity.

Original authors: Aakanksha Tewari, Jai Mangal, Sumit J Darak, Shobha Sundar Ram, Arnav Shukla

Published 2026-04-01

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Aakanksha Tewari, Jai Mangal, Sumit J Darak, Shobha Sundar Ram, Arnav Shukla

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to have a conversation with a friend in a crowded, noisy stadium. You need to shout just loud enough to be heard, but not so loud that you waste energy or annoy everyone else. To do this, you need to know exactly where your friend is standing (location) and how fast they are moving (speed), so you can aim your voice directly at them.

In the world of 6G technology, this "shouting" is communication, and "finding your friend" is radar sensing. The paper you provided describes a brilliant new "smart brain" for a chip that does both jobs simultaneously, but with a twist: it can change its own personality depending on how noisy the stadium is.

Here is the simple breakdown of what the researchers achieved:

1. The Problem: The "Too Slow, Too Loud" Dilemma

Current radar systems are like a person trying to find a friend in a crowd by shouting a list of every single name in the phonebook, one by one.

High Accuracy Mode: They shout very carefully, checking every angle and distance. This finds the friend perfectly, even if the crowd is screaming (noisy). But it takes a long time. While they are shouting, they can't talk to the friend, so the conversation (data transfer) is slow.
Low Accuracy Mode: They shout quickly, guessing the location. This is fast, but if the crowd is too loud, they might shout at the wrong person or miss them entirely.

The old way was to pick one mode and stick with it. If the noise changed, the system got stuck being either too slow or too inaccurate.

2. The Solution: The "Chameleon Chip"

The researchers built a Reconfigurable Radar Signal Processor. Think of this chip as a Chameleon or a Smart Traffic Cop.

The Two Personalities:
1. The Detective (MJARP): When the stadium is noisy (low Signal-to-Clutter ratio), the chip switches to "Detective Mode." It uses a heavy, complex algorithm to carefully sift through the noise, find the exact location, speed, and direction of the target. It's slow but super accurate.
2. The Sprinter (SARP): When the stadium is quiet (high Signal-to-Clutter ratio), the chip switches to "Sprinter Mode." It uses a lightweight, fast algorithm. It doesn't need to check every angle because the signal is clear. It finds the target in a flash.
The Magic Switch: The chip has a built-in sensor that constantly listens to the "noise level." If it hears the crowd getting louder, it instantly morphs into the Detective. If the crowd quiets down, it instantly becomes the Sprinter. It does this without stopping the car or rebooting the system.

3. How It Works (The Hardware)

The chip is built on a Zynq MPSoC, which is like a hybrid car engine. It has two parts working together:

The Brain (ARM Processor): This is the manager. It decides which mode to use based on the noise and manages the overall system.
The Muscle (FPGA): This is the high-speed worker. It does the heavy lifting of calculating distances and speeds.

The researchers optimized the "Muscle" to work with Fixed-Point math. Imagine doing math with a calculator that only has 22 digits instead of a super-precise scientific calculator. It's less precise on paper, but it's much faster and uses less battery. They proved that for this specific job, the "fast calculator" is just as good as the "precise one" but runs 5.6 times faster.

4. The "CLEAN" Algorithm: Removing the Echoes

When you shout in a canyon, you hear echoes. In radar, strong targets (like a big truck) create "echoes" that hide weaker targets (like a bicycle).
The researchers added a CLEAN feature. Imagine the chip is a painter.

It paints the picture of the big truck.
It then "erases" the truck's image from the radar screen.
Now, the hidden bicycle is visible.
It repeats this until all targets are found.
Doing this erasing process directly on the "Muscle" chip (instead of sending data back to the "Brain") saves massive amounts of time.

5. The Results: Why Should We Care?

Speed: This new chip processes radar data 5.6 times faster than the best existing designs.
Efficiency: Because it's faster, the system spends less time "scanning" and more time "talking." This results in a 24% boost in communication speed (throughput).
Battery Life: By using the "fast calculator" (fixed-point) and only using the "heavy detective" mode when absolutely necessary, the chip saves a lot of power.
Versatility: It can handle multiple targets at once (like a whole football team on the field), not just one.

The Big Picture

In the future of 6G, your car or phone needs to "see" the road and "talk" to other cars at the same time. This paper presents a system that doesn't just do both; it does them intelligently. It knows when to be careful and when to be fast, ensuring that your connection stays fast and your safety systems stay accurate, all while saving energy.

It's like having a GPS that doesn't just tell you where you are, but also instantly adjusts its search strategy based on whether you are driving in a quiet suburb or a chaotic city center, ensuring you never get lost and never run out of battery.

1. Problem Statement

The paper addresses the critical challenges in Integrated Sensing and Communication (ISAC) for 6G vehicular networks. ISAC systems must simultaneously perform radar signal processing (RSP) to localize multiple mobile targets (estimating range, azimuth, and Doppler velocity) and communicate with them.

The Bottleneck: Traditional RSP algorithms (like Joint Azimuth-Range Processing, JARP) are computationally intensive and require long processing times. In ISAC, where radar and communication share hardware and spectrum via time-division multiplexing, long RSP latency reduces the time available for data transmission, severely impacting communication throughput.
The Trade-off: High-accuracy algorithms (like JARP) suffer from high latency, while low-latency algorithms often lack accuracy, especially in low Signal-to-Clutter-and-Noise Ratio (SCNR) environments or when resolving multiple targets.
Hardware Limitations: Existing commercial Radar System-on-Chips (SoCs) are primarily designed for sensing only and lack the reconfigurability and integration required for efficient ISAC.

2. Methodology

The authors propose a hardware-software co-designed (HSCD) adaptive architecture implemented on a Zynq Multi-Processor System-on-Chip (MPSoC) (combining an ARM processor and FPGA). The core methodology involves three layers of innovation:

A. Algorithmic Level: Two Complementary Frameworks

The authors developed two distinct RSP frameworks to handle different channel conditions:

Modified Joint Azimuth-Range Processing (MJARP):
- Purpose: High accuracy for low SCNR (noisy) environments.
- Mechanism: Performs full digital beamforming (DBF) and matched filtering (MF) across all search angles and fast-time samples. It uses an iterative CLEAN algorithm to isolate multiple targets by subtracting the strongest target's contribution (including sidelobes) from the signal.
- Optimization: Unlike standard JARP, MJARP optimizes subsequent packets ( $m > 0$ ) by performing selective processing only on the estimated angles and ranges of detected targets, rather than scanning the entire space again.
Sequential Azimuth-Range Processing (SARP):
- Purpose: Low latency for high SCNR (clear) environments.
- Mechanism: A sequential approach that first estimates the angle of arrival (AoA) via non-coherent integration, then estimates range via matched filtering on the selected angle.
- Benefit: Significantly reduces computational complexity by avoiding full 2D scanning for every packet.

B. Adaptive Reconfiguration Strategy

The system dynamically switches between SARP and MJARP based on real-time SCNR sensing.
If the signal is strong (high SCNR), the system uses the fast SARP. If the signal is weak (low SCNR), it switches to the accurate MJARP.
This "on-the-fly" switching maximizes throughput by minimizing processing time without sacrificing localization accuracy.

C. Hardware Architecture & Optimization

Fixed-Point Implementation: To reduce resource usage and power, the authors optimized the algorithms using fixed-point arithmetic (determining optimal word lengths for DBF, FFT, and IFFT blocks) rather than floating-point, achieving up to 40-70% resource reduction.
Super-Resolution Doppler Estimation: Instead of standard FFT for Doppler, the system uses the MUSIC algorithm with a custom hardware IP for Eigenvalue Decomposition (EVD) based on Givens Rotation. This allows for super-resolution (distinguishing targets with very close velocities) which standard FFT cannot achieve.
CLEAN on FPGA: The iterative CLEAN algorithm is implemented directly on the FPGA (Programmable Logic) to remove data transfer overhead between the processor and the accelerator.

3. Key Contributions

Novel Adaptive Architecture: A reconfigurable RSP accelerator that dynamically switches between high-accuracy (MJARP) and low-latency (SARP) modes based on channel conditions, a feature not present in prior ISAC works.
Hardware Innovations:
- Implementation of Givens Rotation-based EVD for MUSIC, offering super-resolution Doppler estimation with lower latency than standard Jacobi methods or Xilinx IP cores.
- Fixed-point optimization that maintains accuracy while drastically reducing FPGA resource utilization (LUTs, DSPs, BRAM) and power consumption.
- Integration of the CLEAN algorithm into the hardware pipeline to handle multiple targets efficiently.
System-Level Integration: The first demonstration of a fully integrated ISAC system on an edge platform (Zynq MPSoC) that handles both sensing and communication, outperforming commercial radar SoCs which are sensing-only.

4. Results

The proposed system was evaluated against state-of-the-art designs (including prior JARP implementations and standard 802.11ad) under various noise conditions and target counts.

Processing Speed: The adaptive architecture achieves up to 5.6× faster RSP processing compared to state-of-the-art JARP designs.
Communication Throughput: By reducing beam alignment time, the ISAC system delivers a 24% improvement in communication throughput compared to MJARP-only systems and up to 88% improvement over standard 802.11ad.
Accuracy:
- Under high SCNR, SARP matches the accuracy of MJARP.
- Under low SCNR, the adaptive switch to MJARP maintains low Root Mean Square Error (RMSE) for range and azimuth.
- The Givens-rotation MUSIC implementation successfully resolves targets with Doppler velocity differences as small as 4 m/s (super-resolution), which standard FFT fails to do.
Hardware Efficiency:
- Power: Up to 26.4% reduction in dynamic power consumption compared to floating-point implementations.
- Resources: Significant reduction in FPGA resource utilization (e.g., ~65% reduction in LUTs for 32-antenna SARP vs. SPFL).

5. Significance

This work provides a critical step toward practical 6G vehicular communication. It solves the fundamental trade-off between sensing accuracy and communication latency in ISAC systems. By demonstrating that an adaptive, fixed-point, reconfigurable architecture on an edge SoC can outperform both commercial radar chips and previous academic designs, the paper establishes a viable path for deploying high-speed, low-latency, and accurate ISAC networks in real-world scenarios. The ability to dynamically adapt to environmental noise ensures robust performance for safety-critical applications like autonomous driving.