Skip to the content
4 layers75 × 44 mmLocalCam1, the printed circuit board in 3D

A small 4-layer board that turns a Luckfox Core1106 system-on-module (Rockchip RV1106G3) into a network camera that never talks to anyone's cloud.

Size
75 × 44 mm
Copper layers
4
Thickness
1.6 mm
Parts
34
Nets
106
Pads
236
Tracks, vias
499, 108
Pours
3
Routed
Yes
KiCad
KiCad 10.0 (file 20260206)
Created
4 Oct 2026
Updated
10 Oct 2026

Overview

LocalCam1 is an open-source printed circuit board by jacksquare914, from the jacksquare914/localcam repository on GitHub. It is a 4-layer board of 75 × 44 mm with 34 parts on 106 nets, routed with 499 tracks and 108 vias.

It is built around the USBLC6-2SC6 and SMAJ6.0CA. Its parts include 9 resistors, 8 capacitors, 3 connectors, 2 ICs and 2 diodes.

Main components

2

README

README.md

LocalCam-1 — a local-only WiFi camera

A small 4-layer board that turns a Luckfox Core1106 system-on-module (Rockchip RV1106G3) into a network camera that never talks to anyone's cloud. A MIPI CSI-2 sensor comes in over an FPC, the module encodes H.264/H.265 on-chip, and the stream is served over RTSP to a local Frigate NVR. Object detection runs on the module's own 1 TOPS NPU, on the camera, not on a server and not on someone else's GPU.

75 × 44 mm, 4 layers, 1.6 mm, built by JLCPCB. USB-C in, microSD for storage, a debug UART header and a recovery button.

Status: not ready to order. 226 of 228 netted pads are connected. Two signals still need a trace and both need a part nudged rather than more routing effort — see What is still open. Everything else has been verified; the evidence is in logs/.


Why this board exists

Every consumer camera in this class wants an account, an app, and an outbound connection. The goal here is the opposite: a camera that is useful precisely because it is isolated. Four requirements drove the whole design.

  1. Nothing leaves the LAN. No cloud service, no phone-home, no vendor account. RTSP to a local NVR and that is all.
  2. Detection happens on the camera. The RV1106G3's NPU runs the model, so there is no video streaming off-box just so a server can look at it.
  3. It must be repairable and legible. A board you cannot read is a board you cannot fix. That drove the routing style (below) as much as any electrical requirement.
  4. One cable. USB-C supplies it; everything else is optional.

What is on it

U1Luckfox Core1106 — RV1106G3, Cortex-A7 + RISC-V MCU + NPU + ISP, 256 MB DDR3L, onboard WiFi and 8 GB eMMC options
J324-pin 0.5 mm FPC to the camera sensor — 2 MIPI data lanes + clock, I²C, MCLK, reset, power-down
J1USB-C 2.0 receptacle — power in and USB device, with CC pull-downs and an ESD part (U2)
J2microSD (Hirose DM3AT), 4-bit SDMMC, with card-detect
H15-pin debug header — UART2 plus the recovery line
S1reset, D2 status LED, F1 1.5 A polyfuse, D1 TVS on the 5 V rail

Full pad-by-pad detail, cross-checked against the vendor pinout, is in docs/pin-assignment-crosscheck.md.


The design decisions worth arguing about

docs/INTENT.md is the long version. The short version:

The stack-up decides everything. JLCPCB JLC04161H-7628: the 0.2104 mm of 7628 prepreg between F.Cu and the ground plane is the only reason the camera pairs can be 100 Ω. If you let the fab substitute a stack-up, the pairs silently stop being 100 Ω. The order settings are printed by hardware/make-fab.sh for exactly that reason.

In1.Cu is a plane, not a routing layer. A MIPI pair's return current flows in the copper directly beneath it; a slot cut in that plane is the classic reason a camera link passes on the bench and drops frames when the enclosure warms up. The router physically cannot place a non-GND track there.

The camera pairs are routed as one trace, then split. A single centreline is routed at the full pair width and both traces are generated as parallel offsets of it, so the gap is constant by construction and the lengths are identical:

pairleg / legintra-pair skew
CSI_D114.51 / 14.51 mm0 µm
CSI_CLK14.72 / 14.72 mm0 µm
CSI_D014.92 / 14.92 mm0 µm

Budget was 100 µm. Inter-pair spread is 410 µm against a 1000 µm budget. No pair changes layer and no pair has a via.

Orthogonal everywhere, with one deliberate exception. Every trace runs parallel or perpendicular to a board edge — 480-odd segments of it — except 18 chamfers, all six of which belong to camera-pair traces. Routed fully orthogonal the clock pair could not stay coupled in the shared FPC channel and came out at 3.0 mm skew, which is a dead MIPI link. Eighteen chamfers in one corridor buy back 0 µm on all three pairs. Everywhere else there is room, so nothing cuts a corner.

The USB-C data pads cross with no vias at all. A 12-pin receptacle interleaves its two D− pads with its two D+ pads at 0.5 mm pitch, so the two shorting links must cross exactly once. A via pair does not fit — a 0.6 mm via needs 0.5 mm of clear radius and the neighbouring pad edge is 0.35 mm away. So one net crosses above the pad row, inside the connector body, and the other crosses below it. They never meet.

Power rails live on In2.Cu, one via per group. 0.5 mm of 1 oz outer copper carries 1.44 A at a 10 °C rise (IPC-2221) and the module draws about 0.6 A. +5V and VBUS_IN vias are doubled, because a single plated barrel carrying a whole rail is the classic warm-attic failure.


Defects found and fixed along the way

These were real, and most were caught by review rather than by the person who made them.

#DefectWhy it mattered
A-7Pad 76 (VCC3V3_RTC) had been tied to +3V3It sits behind a Schottky inside the module; tying it shorts across that diode. Now deliberately unconnected.
A-8Camera I²C pull-ups went to +3V3Pads 16/17 are in the module's 1.8 V IO domain — the vendor pinout confirms it. Pulling them to 3.3 V over-drives 1.8 V pins. Now on +1V8.
A-9VBUS_DET divider was fed from +3V3100k/100k from 3.3 V gives 1.65 V into pad 24 (USB_VBUSDET, a 3.3 V input) — below V_IH, so the pin sat in the indeterminate band. Now from +5V: 2.5 V, a valid high.
A-10/11No decoupling on +1V8; +5V bulk 71 mm from the module~60 mV of droop at 0.6 A with nothing local to absorb a WiFi transmit burst. Added C7 (100 nF on +1V8) and C8 (10 µF local +5V bulk).
A-12Ground pour ran flush to the board edgeIgnored the 2 mm corner fillets. Now inset 0.25 mm, following the radius.
A-13Three vias sat inside padsVia-in-pad wicks solder paste during reflow and is not offered at this price tier.
A-14Two pairs of drills physically overlapped−0.008 and −0.120 mm hole-to-hole against a 0.20 mm minimum. The fab would have rejected it.
A-15U2 (the USB ESD part) was not groundedIts ground pad is fenced in by the USB traces, so the pour could not reach it. It now has its own via to the plane.
A-16The microSD card-detect pins were not in the netlist at allJ2 pins 9/10 had no net, so no connectivity check could report them. The stock Micro_SD_Card symbol has no such pins; J2 now uses Micro_SD_Card_Det2, which matches the DM3AT footprint pin for pin.

A-15 and A-16 are the instructive ones: both were invisible to a net-level check. A net with two pads and one trace looks complete whether or not a third pad is stranded, and a pad with no net cannot be reported as disconnected at all. That is what tools/padaudit.py exists to catch.


Repo layout

text
hardware/     the KiCad 10 project — open LocalCam1.kicad_pro
  lib/        the Core1106 symbol and footprint (not in KiCad's stock libraries)
  make-fab.sh one-shot JLCPCB output generator; run it on a machine with KiCad 10
tools/        the sources of truth and the checkers
  spec.py       NETLIST source of truth — 36 parts, 36 nets. Edit this, never the schematic
  placement.py  PLACEMENT source of truth — part positions and the zone list
  gen.py        spec.py  -> LocalCam1.kicad_sch
  genpcb.py     spec.py + placement.py -> LocalCam1.kicad_pcb (placed, netted, unrouted)
  hroute5.py    the router (orthogonal, with coupled differential pairs)
  hroute6.py    the finisher — routes only what is missing, keeping existing copper
  continuity.py per-net island count + true minimum copper-to-copper gap
  padaudit.py   PER-PAD audit with the ground pour simulated  <- the one that finds real bugs
docs/         design intent, the pinout, the cross-check, routing and mechanical notes
vendor/       Luckfox's own Core1106 pinout spreadsheet, module PDF and footprint package
logs/         verification output, kept as evidence rather than summarised

Building and checking it

The schematic and the board are build outputs. Do not hand-edit them and expect the edit to survive.

bash
cd tools
export KICAD_SYMBOL_DIR=/usr/share/kicad/symbols     # wherever yours live
python3 gen.py        # -> LocalCam1.kicad_sch
python3 genpcb.py     # -> LocalCam1.kicad_pcb  (placed, netted, unrouted)
python3 hroute5.py base.kicad_pcb routed.kicad_pcb   # route it

Then check it with something that did not do the routing:

bash
python3 padaudit.py   ../hardware/LocalCam1.kicad_pcb   # per-pad, pour simulated
python3 continuity.py ../hardware/LocalCam1.kicad_pcb --clearance

padaudit.py walks every pad that carries a net and answers one question: is this pad in the same electrical island as the rest of its net? It simulates the ground pour, because the board ships with pours unfilled so KiCad recomputes them — without that simulation every GND pad looks disconnected and the answer is useless.

Fill the zones in KiCad (press B) and save before plotting, or the pours are not in the Gerbers and the ground plane simply does not exist.

Requires numpy and scipy.

Verified, on the delivered board

text
226 / 228 netted pads connected        (pour simulated)
499 segments: 481 orthogonal, 18 chamfers on camera pairs, 0 other angles
108 vias, all 0.6 / 0.3 mm             (0.15 mm annular ring, min is 0.13)
minimum copper-to-copper gap 0.150 mm  (JLCPCB floor is 0.09)
camera pairs: 0 µm intra-pair skew, 410 µm inter-pair spread
45 of 45 used module pads agree with the vendor pinout

What is still open

Two signals, both placement problems rather than routing ones:

NetPadAtWhat to do
SD_D3J2.2(55.78, 28.51)Six SD signals funnel through one gap between the module and the socket and D3 does not fit. Give it B.Cu with a via each side, or shift J2 about a millimetre.
VBUS_DETR7.2(19.41, 32.60)R7 sits in the channel the USB pair needs between U2 and the module edge. Move it clear — toward y ≈ 35.5 — or rotate it 90°.

The router was given every chance at these: shape search, two-layer maze, two rip-up rounds at widening corridors, and a last-resort pass at 0.15 mm / 0.10 mm. Each attempt either gained nothing or broke three other nets to fix one, and reported that rather than claiming a win.

Open question for review: R7 currently divides +5V (post-fuse). Dividing VBUS_IN (pre-fuse) is what the pin name literally means and still reads high if the polyfuse opens; +5V instead answers "is my rail actually up?". Both give 2.5 V and both are valid. One line in spec.py.

Reviewing this board — where to look first

If you have ten minutes: docs/INTENT.md §4 for the arguable decisions, then logs/VERIFY-pads.log for what is actually connected. If you have an hour, run padaudit.py yourself and disagree with it.

Things I would most like a second opinion on:

  • The two open nets above — is nudging R7 and J2 the right call, or should the floorplan change more deeply?
  • The USB pair is not coupled between U1 and U2, because U2's +5V pin sits physically between D− and D+. Over ~8 mm at USB 2.0 that should be fine. Second opinion welcome.
  • +5V reaches the module through a ~71 mm path. C8 was added locally to cover it, but a different floorplan would shorten the path instead.
  • Whether grounding both card-detect pins is the right choice versus bringing one out to a GPIO with a pull-up.

Sources

Licence

No licence has been chosen yet — which by default means all rights reserved, so pick one before sharing this widely.

Parts

34
QtyValueFootprintReferences
4100nFCapacitor_SMD:C_0402_1005MetricC2, C4, C6, C7
4M2MountingHole:MountingHole_2.2mm_M2MH1, MH2, MH3, MH4
310uF/16VCapacitor_SMD:C_0805_2012MetricC3, C5, C8
2100kResistor_SMD:R_0402_1005MetricR7, R8
24.7kResistor_SMD:R_0402_1005MetricR5, R6
25.1kResistor_SMD:R_0402_1005MetricR1, R2
1+3V3TestPoint:TestPoint_Pad_D1.5mmTP2
1+5VTestPoint:TestPoint_Pad_D1.5mmTP1
11.5A holdResistor_SMD:R_1206_3216MetricF1
110kResistor_SMD:R_0402_1005MetricR3
11kResistor_SMD:R_0402_1005MetricR9
1330RResistor_SMD:R_0402_1005MetricR4
147uF/16VCapacitor_SMD:C_0805_2012MetricC1
1Core1106 (RV1106G3)Core1106:Core1106-SMTU1
1DEBUG UARTConnector_PinHeader_2.54mm:PinHeader_1x05_P2.54mm_VerticalH1
1FPC 0.5mm 24p to SC3336Connector_FFC-FPC:Hirose_FH12-24S-0.5SH_1x24-1MP_P0.50mm_HorizontalJ3
1GNDTestPoint:TestPoint_Pad_D1.5mmTP3
1RESETButton_Switch_SMD:SW_Push_1P1T_NO_CK_KMR2S1
1SMAJ6.0CADiode_SMD:D_SMAD1
1STATUS red Vf~2.0VLED_SMD:LED_0603_1608MetricD2
1USB-C 2.0Connector_USB:USB_C_Receptacle_HRO_TYPE-C-31-M-12J1
1USBLC6-2SC6Package_TO_SOT_SMD:SOT-23-6U2
1microSDConnector_Card:microSD_HC_Hirose_DM3AT-SF-PEJM5J2

Similar boards

  • 4 layers95 × 58 mmEsp32Tap, the printed circuit board in 3D

    Esp32Tap

    scottmsilver/treddy

    Controlling my procor-9.3x treadmill

    94 parts2
  • 2 layers169.4 × 48.8 mmUsb morse keyer, the printed circuit board in 3D

    Usb morse keyer

    dilshan/usb-morse-keyer

    PIC16F886 based CW auto key controller with USB support

    84 parts17MIT
  • 2 layers68 × 81.5 mmController, the printed circuit board in 3D

    Controller

    orlopau/coffee-scale

    Scale for espresso and filter coffee.

    55 parts11
  • 4 layers260 × 110 mmPandore, the printed circuit board in 3D

    Pandore

    filliformes/pandore

    pandore is an open-source digital audio instrument protyping platform based on the LattePanda Mu.

    804 parts2
  • 4 layers208 × 129 mmMainboard v2 r1, the printed circuit board in 3D

    Mainboard v2 r1

    Benjamin-girard-1/geophys_acquisition_module

    Modular ESP32-S3 geophysical data-acquisition system with custom hardware, multi-channel ADC, GNSS, IMU, removable analog sensor cards, firmware, and host tools

    287 parts
  • 2 layers14 × 17 mmME6206A33XG, the printed circuit board in 3D

    ME6206A33XG

    lessthan00/ME6206A33XG

    kicad LDO SOT-23-3

    8 parts
  • 4 layers80 × 80 mmAuxio board, the printed circuit board in 3D

    Auxio board

    MatthiasGeorgImhof/Hardware_CSAT

    Hardware design files for Cubesat PCBs

    129 parts
  • 4 layers94.5 × 69.7 mmDemo LoRa Node, the printed circuit board in 3D

    Demo LoRa Node

    nborkowetz/STM32WL_LoRaSensorNodeDemo

    STM32WL55-based low-power LoRaWAN platform for testing modular sensors and end-to-end IoT infrastructure.

    109 parts

Questions

How big is the LocalCam1 PCB?

It measures 75 × 44 mm and has 4 copper layers, 1.6 mm thick.

What parts does the LocalCam1 use?

34 parts, built around the USBLC6-2SC6 and SMAJ6.0CA. 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 LocalCam1 KiCad files?

From the jacksquare914/localcam repository on GitHub, at the commit this page was made from: the board is hardware/LocalCam1.kicad_pcb and its schematic hardware/LocalCam1.kicad_sch. PCBKit opens the board in the browser, with nothing to install.

Can I use the LocalCam1 design in my own project?

Its repository states no licence, so its author keeps all rights: ask jacksquare914 before reusing the design.