Skip to the content
4 layers58 × 56.1 mmRadar board, the printed circuit board in 3D

A low-cost launch monitor that recovers a ball's launch vector from one high-speed monochrome camera and one 24 GHz CW Doppler radar — the RF front-end, the acquisition firmware, the synchronized recorder, and the…

Size
58 × 56.1 mm
Copper layers
4
Thickness
1.61 mm
Parts
82
Nets
45
Pads
226
Tracks, vias
685, 71
Pours
3
Routed
Yes
KiCad
KiCad 10.0 (file 20260206)
Created
3 Sep 2026
Updated
3 Sep 2026

Overview

Radar board is an open-source printed circuit board by wzwqoo, from the wzwqoo/single_camera_single_radar_rx_launch_monitor repository on GitHub. It is a 4-layer board of 58 × 56.1 mm with 82 parts on 45 nets, routed with 685 tracks and 71 vias.

It is built around the TLV9001IDCK and TPS74801DRCR. Its parts include 39 capacitors, 33 resistors, 4 connectors and 3 ICs.

Main components

2

README

README.md

Single-Camera / Single-Radar Golf Launch Monitor

A low-cost launch monitor that recovers a ball's launch vector from one high-speed monochrome camera and one 24 GHz CW Doppler radar — the RF front-end, the acquisition firmware, the synchronized recorder, and the processing that turns both streams into a single answer.

Monocular trajectory fit

Left: per-frame ball detections tracked through the image. Right: the same flight reconstructed in the world frame. Header: the resulting launch state — speed, azimuth, elevation.

Why two sensors

Each sensor alone measures half the problem, and each is blind exactly where the other is strong:

measures wellcannot measure
Radar (CW Doppler)speed along the line of sight, to ≈0.5 mphlaunch direction — one channel gives one scalar
Camera (mono, 210 fps)launch azimuth and elevation, over-determined by ~10–15 framesabsolute depth — a single view is nearly blind along its own optical axis

So the system is built around a deliberate split of labour: radar supplies the speed magnitude, the camera supplies the direction. The camera is never asked for absolute scale, and the radar is never asked for geometry.

mermaid
flowchart LR
    A["OV9281 camera<br/>640×400 @ 210 fps"] --> C["blob / contour detection<br/>→ Kalman track"]
    C --> D["bundle adjustment<br/>vs. known wall plane"]
    D --> F["launch azimuth<br/>+ elevation"]
    B["K-LC7 24 GHz radar<br/>→ STM32 → USB"] --> E["Doppler STFT<br/>→ peak speed"]
    E --> G["line-of-sight<br/>speed |v|"]
    F --> H["fused launch vector<br/>v / cos γ, az, el"]
    G --> H

A single tool records both sensors together, so every radar clip has a matching optical trajectory and vice versa.

Hardware

Both sensors are commodity parts; the only custom electronics is the radar's IF carrier board. Full details in hardware/ and firmware/.

Camera

OV9281 camera module

The camera as fielded: the OV9281 sensor and lens on a flex ribbon, with the USB interface board at the far end.

SensorOmniVision OV9281 — 1 MP monochrome CMOS, global shutter, ≈$15 module
Mode640×400 MONO8 at 210 fps, hardware MJPEG over USB 2.0
Opticswide-angle, HFOV ≈ 75° ⇒ f ≈ 417 px
Geometrycamera 53 ft from a wall plane; a known wall circle spans ≈27 px, confirming the focal length

Global shutter is a correctness requirement rather than a convenience: apparent ball diameter is the depth measurement, and rolling-shutter skew would corrupt the very quantity being calibrated. Monochrome readout drops the Bayer mosaic, which matters when the ball is only 3–25 px across.

Radar

K-LC7 module on the custom IF carrier board

The K-LC7 transceiver module (top, patch-array antenna side) mounted on the custom IF carrier board carrying the two-stage amplifier chain.

ModuleRFbeam K-LC7, 24.0–24.25 GHz ISM, CW Doppler transceiver with integrated VCO
Receivetwo coherent I/Q pairs; I/Q phase 80–100°, amplitude imbalance ±3 dB
Sensitivity−96 dBm (1 kHz BW, 6 dB S/N)
IF boardcustom carrier: two ×32 stages = 60 dB, high-pass corner ≈3.4 kHz
MCUSTM32F446 — dual ADC + DMA at 24 kSPS/channel, FreeRTOS, USB CDC
Antenna study24.125 GHz stepped horn, 65×50 mm aperture, modelled in MATLAB

The module's 100 kHz VCO modulation bandwidth is what forces the CW architecture: FMCW at sports-ball speeds would need an unambiguous velocity far above the ≈28 mph this bandwidth allows. CW buys speed and spin at the cost of all range information.

Repository layout

Each folder has its own README with the details.

PathContents
hardware/K-LC7 IF carrier board (KiCad project + fab/ Gerbers), 24 GHz horn-antenna study, module datasheet
firmware/STM32F446 acquisition firmware — timer-paced dual-ADC DMA, on-board Doppler trigger, USB CDC raw streaming
capture/Synchronized camera + radar recording, camera calibration, clip trimming
tracking/Camera pipeline: detection → mono track → 3-D trajectory fit → launch angles
radar/Doppler speed and micro-Doppler spin-rate extraction from recorded I/Q
data/radar_captures/Eight raw throw captures, with the speed measured from each
docs/Figures used by these READMEs

Getting a measurement

bash
pip install -r requirements.txt

# 1. Calibrate the camera (once, per lens)
python capture/1_calibrate_camera.py

# 2. Record camera + radar together; space bar starts/pauses both streams
python capture/3_record_combined.py /dev/ttyACM0

# 3. Trim the radar clips to the frames you kept
python capture/6_trim_to_frames.py --frames camera_frames --radar radar_data

# 4. Radar → speed magnitude
python radar/speed_test.py radar_data/rx2_YYYYMMDD_HHMMSS.npz

# 5. Camera → launch direction (and the full reconstruction)
python tracking/explore.py camera_frames --out throw01

Firmware builds with STM32CubeIDE or the included CMake preset; the horn model needs MATLAB with the Antenna Toolbox.

Limitations

This is a research build, and the honest scope is narrower than "launch monitor" suggests:

  • The camera cannot measure speed. Depth from apparent ball size is both noisy and biased — a small fast ball is systematically over-measured, so the size-to-depth channel reads too shallow. A single calibration constant absorbs that bias, but it was fitted against only four radar-validated clips (≈5 mph MAE). The camera's standalone absolute-speed estimate should not be trusted; only its direction output is well-posed.
  • A known wall plane is required. The reconstruction becomes well-posed only when the wall supplies an absolute-depth datum. Take the wall away — an open driving range — and the geometry degenerates again.
  • The IF high-pass eats slow shots. The board's single-pole ≈3.4 kHz corner passes Doppler above ≈47 mph. Every recorded throw was 31–46 mph, i.e. below the corner, so carrier and sidebands both needed a +3–5 dB insertion-loss correction. The corner belongs an octave lower.
  • CW gives no range. With no range gate there is no range-based clutter rejection, and the target must be isolated by Doppler alone.
  • Spin rate yes, spin axis no. Spin rate is recoverable from micro-Doppler sideband spacing at ≈10.6 dB SNR. The spin axis is not: it depends on sideband amplitude through a calibration constant with no ground truth available, and the ≈21% per-throw amplitude error plus a +20–30% max-of-N selection bias propagate to tens of degrees — saturating worst near the pure-backspin geometry that matters most. A motorized spin machine with known rate and settable axis is a prerequisite for going further, and no spin-axis code ships here.
  • Validated on baseballs, not golf balls. The dataset is eight thrown-baseball clips at 31–46 mph — an accessible stand-in with the same launch phenomenology, but an order of magnitude below real golf launch speeds.
  • Offline, not real-time. explore.py post-processes a complete clip; the question it answers is "given every frame, how trustworthy is the answer", not "how fast can this converge".
  • A hardware failure is on the record. Two IF channels collapsed to a DC short one after another, consistent with over-stress into unprotected high-impedance op-amp inputs. Input protection and grounding are the first thing a successor board should fix.

Parts

82
QtyValueFootprintReferences
14100nFCapacitor_SMD:C_0805_2012MetricC1, C10, C11, C14, C16, C18, C25, C33, C34, C37, C39, C42, C44, C45
1110nFCapacitor_SMD:C_0805_2012MetricC9, C12, C13, C15, C17, C19, C20, C21, C22, C26, C40
931K2Resistor_SMD:R_0805_2012MetricR14, R15, R16, R17, R18, R19, R20, R21, R34
81KResistor_SMD:R_0805_2012MetricR22, R23, R24, R25, R26, R27, R28, R29
8330pFCapacitor_SMD:C_0805_2012MetricC23, C24, C27, C28, C29, C30, C31, C32
4100Resistor_SMD:R_0805_2012MetricR2, R3, R4, R5
4100KResistor_SMD:R_0805_2012MetricR10, R11, R12, R13
422uFCapacitor_SMD:C_0805_2012MetricC35, C36, C41, C43
44K7Resistor_SMD:R_0805_2012MetricR6, R7, R8, R9
310KResistor_SMD:R_0805_2012MetricR30, R32, R33
210uFCapacitor_SMD:C_0805_2012MetricC2, C38
2600Ω@100MHzInductor_SMD:L_0805_2012MetricFB1, FB2
2OPA4354AIPWRPackage_SO:TSSOP-14-1EP_4.4x5mm_P0.65mmU1, U2
122Resistor_SMD:R_0805_2012MetricR31
1Conn_01x03_PinConnector_PinHeader_2.54mm:PinHeader_1x03_P2.54mm_VerticalJ5
1Conn_01x03_SocketConnector_PinSocket_2.54mm:PinSocket_1x03_P2.54mm_VerticalJ1
1Conn_01x04_PinConnector_PinHeader_2.54mm:PinHeader_1x04_P2.54mm_VerticalJ4
1Conn_01x06_SocketConnector_PinSocket_2.54mm:PinSocket_1x06_P2.54mm_VerticalJ2
1TLV9001IDCKPackage_TO_SOT_SMD:SOT-353_SC-70-5U3
1TPS74801DRCRPackage_SON:VSON-10-1EP_3x3mm_P0.5mm_EP1.65x2.4mmTPS74801DRCR1

Similar boards

  • 2 layers10.5 × 12 mmESP32 USB spy, the printed circuit board in 3D

    ESP32 USB spy

    cpla33/ESP32_USB-spy

    Smallest ESP32-based board ever possible

    21 parts
  • 4 layers36.6 × 100 mmEchometer, the printed circuit board in 3D

    Echometer

    wzwqoo/Echometer

    KiCad 10 project (v4.kicadpro, schematic + 4-layer PCB), fabrication outputs in fab/, BOM in v4.csv, and the parametric acoustic-horn generator ultrasonichornfreecad.py.

    78 parts
  • 4 layers86 × 55 mmDevice, the printed circuit board in 3D

    Device

    jayis1/hacker-devices

    Full-cycle hardware designs for security research devices — each device is a subfolder with its own README, schematics, PCB layout, and software stack

    22 parts2
  • 4 layers110 × 64 mmSYCL Badge 2024, the printed circuit board in 3D

    SYCL Badge 2024

    ZigEmbeddedGroup/sycl-badge

    The badge for the SYCL conference! https://zigembeddedgroup.github.io/sycl-badge/

    145 parts39
  • 4 layers101.5 × 44.8 mmShori, the printed circuit board in 3D

    Shori

    ArnavMK/WaveForge

    A fully custom Solid State Hybrid Guitar Amplifier

    135 parts4
  • 4 layers94.3 × 81.8 mmYukaLink revA, the printed circuit board in 3D

    YukaLink revA

    Misaka0x2730/YukaLink

    Dual targets USB+ETH ARM JTAG/SWD debugger (BlackMagic port) based on RP2350. JTAG and SWD interfaces are implemented on PIO.

    321 parts3GPL-3.0
  • 4 layers85 × 56 mmRadiohead, the printed circuit board in 3D

    Radiohead

    mpvano/radiohat2

    Software and Documentation for the Radiohat2 Project

    175 parts2MIT
  • 2 layers46.7 × 20.3 mmSoundkey, the printed circuit board in 3D

    Soundkey

    daimapi/soundkey

    38 parts1

Questions

How big is the Radar board PCB?

It measures 58 × 56.1 mm and has 4 copper layers, 1.61 mm thick.

What parts does the Radar board use?

82 parts, built around the TLV9001IDCK and TPS74801DRCR. The full list of its parts is on this page, with links to each part's symbol and footprint where PCBKit's library has them.

Where can I download the Radar board KiCad files?

From the wzwqoo/single_camera_single_radar_rx_launch_monitor repository on GitHub, at the commit this page was made from: the board is hardware/radar_board/hardware.kicad_pcb and its schematic hardware/radar_board/baseband.kicad_sch. PCBKit opens the board in the browser, with nothing to install.

Can I use the Radar board design in my own project?

Its repository states no licence, so its author keeps all rights: ask wzwqoo before reusing the design.