The problem
As more vehicles carry several radars in the same 76–81 GHz band, one radar's chirps can fall inside another's receiver bandwidth. Interference appears as short bursts in the beat signal that spread across the spectrum, raising the noise floor and hiding weak targets such as pedestrians, or creating ghost targets. Mitigation must work in real time without degrading detection of real objects.
Objectives
- Model victim and aggressor FMCW radars with configurable chirp parameters and geometry.
- Characterise interference effects on the beat signal and range–Doppler map.
- Implement time-domain, interpolation, waveform-diversity and learned mitigation methods.
- Compare methods on SINR improvement, detection probability, false alarms and cost.
- Evaluate under multi-aggressor traffic scenarios.
What you need
Tools and toolboxes
- MATLAB — Signal simulation, mitigation algorithms and Monte Carlo evaluation.
- Radar Toolbox — Automotive radar waveform design and scenario-based radar simulation.
- Phased Array System Toolbox — FMCW waveforms, propagation and CFAR detection objects.
- Automated Driving Toolbox — Driving Scenario Designer to place victim and aggressor vehicles in traffic.
- Deep Learning Toolbox — Training a convolutional denoiser on interfered beat signals or range–Doppler maps.
Software
- MATLAB with Radar Toolbox, Phased Array System Toolbox, Automated Driving Toolbox, Signal Processing Toolbox and Deep Learning Toolbox
Extra parts and data
- A GPU (optional) — Speeds up training of the learned denoiser; the signal processing methods run on a CPU.
Key concepts
- Mutual interference
- When an aggressor radar's chirp sweeps through the victim's receiver band, the victim sees a short, high-amplitude chirp-like burst in its beat signal.
- SINR
- Signal-to-interference-plus-noise ratio. Interference raises the noise floor, lowering SINR and the probability of detecting weak targets.
- Time-domain zeroing
- Detecting interference samples by amplitude and setting them to zero before the FFT. Simple, but it removes signal too and raises sidelobes.
- Waveform randomisation
- Varying chirp start frequency, slope or timing from frame to frame so interference does not align coherently and can be averaged out.
Step-by-step guide
- 1
Frame the questions and scenarios
State the research questions: how much does interference degrade detection of weak targets in realistic traffic, and which mitigation gives the best detection with acceptable cost? Define scenarios: one aggressor head-on, an aggressor in the adjacent lane, several aggressors in dense traffic, aggressors with similar and different chirp slopes. Define metrics: SINR gain, detection probability for a pedestrian-like target, false alarms per frame and processing time.
Checkpoint: Research questions, a scenario matrix and metric definitions.
- 2
Model victim and aggressor radars
Model the victim as a 77 GHz FMCW radar with defined chirp parameters and sampling. Model aggressors with independent parameters, random start times and their own transmit power. For each victim chirp, compute the aggressor signal arriving at the victim receiver with one-way propagation loss, mix it with the victim chirp and pass it through the receiver's anti-aliasing filter, so only the part inside the IF bandwidth remains.
Checkpoint: Simulated beat signals showing interference bursts at the expected times.
- 3
Characterise interference effects
Compare range–Doppler maps with and without interference for a scenario with a car and a pedestrian-like target. Measure the noise floor increase, SINR loss and missed detections for the weak target, and look for ghost targets when aggressor and victim chirps have similar slopes. Vary aggressor distance, slope difference and number of aggressors and plot noise floor rise against each.
Checkpoint: Plots of noise floor and SINR loss against interference parameters, with example maps.
- 4
Place radars in traffic scenarios
Use the Driving Scenario Designer to build a multi-lane road with oncoming and same-direction traffic, giving several vehicles a front radar. At each time step, compute geometry between the victim and every aggressor that falls within each other's field of view and feed it to the signal model. This gives realistic interference statistics instead of hand-picked cases. Log, for every interference event, the aggressor's relative position and chirp parameters so results can later be grouped by geometry.
Checkpoint: Scenario runs producing interference events with counts and strengths over time.
- 5
Implement time-domain methods
Detect interference samples with a robust amplitude threshold (for example a multiple of the median absolute deviation) and implement zeroing, windowed zeroing with a smooth taper, and autoregressive or linear interpolation across the corrupted segment. Measure SINR gain and detection probability for each, and show the sidelobe effects introduced by zeroing long segments. Check how each method behaves when interference covers a large fraction of the chirp, because methods that work for short bursts can fail badly when several aggressors overlap.
Checkpoint: SINR gain and detection results for each time-domain method across scenarios.
- 6
Implement waveform diversity
Randomise the victim's chirp start frequency or slope within limits from chirp to chirp or frame to frame and compensate in processing so targets remain coherent. Measure how interference energy spreads instead of forming ghosts, and how much SINR is recovered after Doppler integration. Report any loss in range or velocity accuracy caused by randomisation. Keep the randomisation within limits that preserve the required range resolution and maximum range from the victim's specification.
Checkpoint: Ghost suppression and SINR results for randomised waveforms, with accuracy impact.
- 7
Train a learned denoiser
Generate a large dataset of paired clean and interfered beat signals or range–Doppler maps from the simulator with randomised scenarios, splitting by scenario seed into training, validation and test sets. Train a convolutional network with Deep Learning Toolbox to predict the clean signal. Evaluate SINR gain and detection on test scenarios and on interference parameters outside the training range to check generalisation.
Checkpoint: A trained denoiser with in-range and out-of-range test results.
- 8
Compare all methods
Run Monte Carlo trials over the full scenario matrix for no mitigation, each time-domain method, waveform diversity, the learned denoiser and promising combinations. Report SINR gain, detection probability for the weak target, false alarms and processing time per frame. Use confidence intervals and identify which method is best in which situation rather than claiming one overall winner. Include a clean, interference-free reference in every table so readers can see how close each method gets to the ideal.
Checkpoint: A comparison table with confidence intervals and a situation-by-situation recommendation.
- 9
Assess real-time feasibility
Estimate the computational cost of each method in operations per frame and measure MATLAB runtime, then relate it to a radar frame time of around 50 ms on a typical embedded radar processor. For the learned denoiser, count parameters and multiply–accumulate operations. Discuss whether each method fits the budget and what simplifications would be needed. Note which steps can run on a radar processor's hardware accelerators, such as FFT engines, and which would need the general-purpose core.
Checkpoint: A cost table relating each method to a frame-time budget.
- 10
Write the thesis chapter
Present the interference model, characterisation, traffic-scenario statistics, each mitigation method, the comparison and the feasibility analysis. State the limitations: simulated interference only, simplified propagation without multipath and no hardware validation. Discuss how results could be checked with recorded radar data containing interference and how standardised coordination could complement signal processing. Release the simulator configuration and the generated datasets, with the seeds used, so the learned and classical methods can be compared by others under identical conditions.
Checkpoint: A complete chapter with reproducible simulation scripts and datasets.
Testing & validation
- Noise floor rise and SINR loss against interference parameters.
- SINR gain, detection probability and false alarms per method with confidence intervals.
- Generalisation of the learned denoiser outside its training range.
- Processing cost against a frame-time budget.
Expected results
A configurable radar interference simulator tied to traffic scenarios, a characterisation of interference effects on weak-target detection and a fair, quantified comparison of mitigation methods with real-time feasibility.
Research angle for your PhD
Research questions: how does mutual interference in realistic traffic affect weak-target detection, and which mitigation, or combination, gives the best detection at an embedded processing cost? Metrics: SINR gain, detection probability, false alarms, processing cost, generalisation outside training conditions for learned methods. Baselines: no mitigation and zeroing. Scenario-driven interference statistics and an honest generalisation test for learned denoisers distinguish the work and make it suitable for radar journals.
Take it further
- Validate with recorded radar data containing real interference.
- Combine detection-and-avoid strategies that change the victim's chirp when interference is sensed.
- Study interference between corner radars in parking scenarios.
- Extend the learned denoiser to the full range–Doppler–angle cube.
Viva and review questions
Why does interference appear as a short burst in the beat signal?
The aggressor's chirp only falls inside the victim's IF bandwidth for the short time when the two chirp frequencies are close, so only a few samples are affected.
Why does zeroing raise sidelobes?
Setting samples to zero acts like multiplying by a gapped window, which spreads energy in the spectrum; tapering or interpolation reduces this.
When do ghost targets appear?
When aggressor and victim chirps have similar slopes, the interference produces a near-constant beat frequency that looks like a real target.
How did you test whether the learned denoiser generalises?
By evaluating it on interference parameters outside the training range, such as slope differences and numbers of aggressors not seen in training.
Learn it at DIYguru
These DIYguru courses and programmes teach what this project needs.