The radar model#

Why the model is the algorithm context#

Range, Doppler, and angle are not properties of an ndarray. They depend on chirp slope, sampling rate, emission timing, wavelength, and physical antenna phase centres. Radar is therefore an immutable aggregate rather than a loose bag of optional keyword arguments.

Component	Responsibility
`FMCW`	sawtooth chirp timing and frequency law
`Sampler`	ADC shape, rate, bit depth, and frame period
`Transceivers`	TX/RX phase-centre coordinates in metres
`MIMOScheme`	channel map, emission order, decoding, timing correction
`Calibration`	ADC-, range-, and array-domain correction arrays
`ProcessingConfig`	layered algorithm defaults and retained products

Instances are frozen dataclasses. Shape and physical consistency are validated at construction, including ADC capture duration versus the chirp ramp and TX/RX counts versus the MIMO strategy.

Derived quantities#

The following properties are calculated from the physical model:

rangeResolution, rangeBinSize, and maxUnambiguousRange
velocityResolution, velocityBinSize, and maxUnambiguousVelocity
virtualArray, duplicatePhaseCenters, and arrayGeometry
unambiguousFov, rangeAxis, velocityAxis, azimuthAxis, and elevationAxis

Resolution and FFT-bin spacing are intentionally distinct. Zero padding can make rangeBinSize smaller, but cannot improve rangeResolution.

Hardware and capture profiles#

TI2243CascadeRadar represents the four-chip AWR2243 cascade. RaDelft, ColoRadar, and ColoRadar+ are capture profiles that instantiate that hardware with different waveform, sampler, TX order, and calibration metadata:

from pyradar.base import TI2243CascadeRadar

radar = TI2243CascadeRadar.radelft()

AWR1843Radar.rampcnn() provides the RAMPCNN/CRUW profile. Dataset readers may replace fallback profile values with parsed JSON, MAT, or text metadata. A profile must not force a desired bin size by changing physical parameters.

Processing configuration#

Configuration is split by stage:

from pyradar.base import (
    CFARConfig,
    DoAConfig,
    FFTConfig,
    ProcessingConfig,
)

processing = ProcessingConfig(
    fft=FFTConfig(
        rangeFftSize=512,
        dopplerFftSize=128,
        rangeWindow="hann",
        dopplerWindow="hann",
    ),
    cfar=CFARConfig(method="os", pfa=1e-4),
    doa=DoAConfig(method="auto"),
    retainRangeAngle=True,
)

The default pipeline is detection-first. Dense range-angle and range-Doppler-angle cubes are opt-in because a 4D cascade cube can otherwise consume several gigabytes.

Data products#

ADCFrame and RadarCube attach names to dimensions. DetectionSet, PointCloud, ClusterSet, TrackSet, and FrameResult make downstream fields explicit and preserve source bins. These contracts are preferable to ndarray columns whose meanings change between scripts.