Solution · Smart Vision

Smart Vision Solution
From Optical Selection to Mass Production

A camera product is not decided by the NPU number. It is decided by whether the lens matches the sensor, whether the ISP is tuned on your scenes, whether the model still holds accuracy after INT8 quantization, and whether the sealed enclosure keeps the sensor cool enough at night. BesTom covers that whole chain on Rockchip vision platforms — RV1106B, RV1126, RV1126B, RK3576 and RK3588.

AI-ISP 2.0 / 12M AI-ISP 0.5 – 6 TOPS NPU MIPI CSI-2 multi-camera RKNN INT8 quantization Battery-powered IPC IP66 / IP67 enclosure
Use cases

Where it fits

The same design discipline serves a 2 MP battery camera and a multi-channel analytics box — what changes is the platform tier, not the method.

Network IPC (bullet / dome / turret)

Standalone camera with on-board encoding and RTSP/ONVIF output. The mainstream volume product for this platform family.

Face recognition & access control

Local verification on the device, no cloud round-trip. Liveness check and lighting conditions drive the sensor and ISP choice.

People counting & footfall analytics

Overhead counting with cross-line tracking. BesTom already runs this application in production — see the people-counting case.

Behaviour & safety event analysis

Intrusion, fall, PPE compliance, restricted-area entry. Requires tuned detection plus a false-alarm review loop, not just a model.

Industrial machine vision

Presence/absence, positioning, surface defect. Fixed lighting and fixed optics — repeatability matters more than resolution.

Vehicle & off-highway vision

Wide-temperature operation, vibration, reverse polarity and load-dump tolerance. Enclosure and power design dominate here.

Agricultural & ecological observation

Long unattended runtime, remote sites, solar or battery supply. Duty-cycle design is the whole engineering problem.

Battery-powered vision node

PIR-triggered wake, seconds-level capture and upload, months of standby. RV1106B with SiP DDR exists for exactly this shape.

Platform selection

Which Rockchip vision tier fits your product

Specs below follow the same platform matrix published on our Rockchip SoM page. Final parameters are governed by the official Rockchip datasheet for the exact part number.

TierPlatformNPUISP / videoKey interfaceChoose it when
Entry · batteryRV1106B
Cortex-A7 + MCU, SiP DDR 512 Mb / 1 Gb / 2 Gb
0.5 TOPS8 MP ISP · 4K@25 encodeMIPI-CSI, RMII+PHY, USB 2.0Standby power and BOM are the binding constraints; SiP DDR removes external memory and its power/area cost
Entry · costRV1106 / RV1103
Single Cortex-A7
0.5 TOPSLightweight visionStreamlinedCheapest possible vision node and the algorithm is genuinely light
MainstreamRV1126
4× Cortex-A7
2 TOPSAI-ISP 2.0, black-light full colourMIPI-CSI, EthernetYou need full-colour night imaging and a mature proven mass-production stack
Mainstream · upRV1126B
4× Cortex-A53
3 TOPS12 MP AI-ISP · 4K@30 decode / 4K@45 encodeMIPI-CSI ×2 (4×2-lane), RGMII+PHY, USB 3.0Dual camera, higher resolution, or you need USB 3.0 and external DDR flexibility
Multi-channelRK3576
4× Cortex-A72 + 4× Cortex-A53
6 TOPS (Transformer)4K@120 decode / 4K@60 encodePCIe, multi-displaySeveral cameras at once, or Transformer-class models on the device
FlagshipRK3588
4× Cortex-A76 + 4× Cortex-A55
6 TOPS8K@60 decode / 8K@30 encodePCIe 2.1 ×4, 4× SATA, dual GbENVR-with-analytics, many channels, or heavy local storage expansion
Reference architecture

The vision chain, and what actually bites

Every block below has a failure mode that only shows up in the field. We design against those, not around them.

SubsystemWhat we specifyThe failure mode we design against
OpticsLens selection, CRA match, IR-cut, depth of field, distortion budgetLens CRA mismatched to the sensor micro-lens array shows up as edge colour shift and vignetting — accepted by MTF alone. We verify relative illumination too.
Image sensorOptical format, pixel size, shutter type, dynamic rangeChasing megapixels instead of pixel size makes low-light performance worse. Moving subjects need global shutter regardless of resolution.
ISPAI-ISP 2.0 / 12 MP AI-ISP, 3DNR, wide dynamic range, defogTuning that passes a colour chart in the lab but degrades on real scenes at dusk or under mixed indoor lighting.
NPU & model0.5 – 6 TOPS, INT8 quantization, RKNN deploymentQuantization silently drops accuracy. Frame rate looks great, recall is gone. We require a before/after accuracy report, not just FPS.
EncodingH.264/H.265, CBR/VBR, ROI-based smart encodingSmart encoding reduces bitrate in daylight, then night noise pushes it back up and saturates the uplink.
StorageeMMC / SD / NVR backhaul, circular recordingWrite amplification and bad-block exhaustion on continuously overwritten media.
NetworkRMII/RGMII + integrated PHY, PoE budgetThe PoE budget is signed off without the IR illuminator peak. The device browns out the moment it gets dark.
PowerLow-power rails, sleep/wake strategy, duty cyclingAverage battery current is set by duty cycle; PIR false triggers and wake settling time dominate the real consumption.
EnclosureIP66/IP67 sealing, anti-fog, thermal pathA sealed housing traps sensor heat, raising dark current and noise. Thermal and noise behaviour must be validated together.
Deliverables

What you actually receive

Each stage closes on a measurable acceptance criterion — not on a slide.

StageDeliverableAcceptance criterion
FeasibilityOptics and platform selection report with BOM range and alternatesReal sample frames from your target scene, plus measured frame rate and power
HardwareSchematic, PCB, EMC/ESD design reviewCritical-signal integrity review and power-rail ripple within budget
Driver / BSPBSP, sensor driver, ISP parameter setPasses capture under every lighting condition in the agreed scene list
AlgorithmModel quantization, RKNN deployment, accuracy reportPost-quantization metric drop within the agreed threshold
Mechanical3D model, thermal and sealing designIP rating test report
Mass productionTest fixture and SOP, yield ramp supportDarkroom, MTF, colour reproduction and burn-in gates in the line
Why BesTom

Why BesTom

Design house since 2008 — a design service, not a storefront. Finished-product volume runs through MeToclaw, the product brand of the same group.

Vision applications already in production

People counting on RK3588/RK3576 and commercial-IoT vision on RV1126/RV1106 — both documented as client cases, not concepts.

ISP tuning in-house

Image quality is decided by the tuning loop, so it stays inside the team instead of being outsourced to a module vendor.

Platform range under one roof

From a 0.5 TOPS battery camera to a 6 TOPS multi-channel analytics box — the same team can move your product up or down a tier.

Shenzhen prototype-to-volume chain

Rapid prototyping plus contracted manufacturing, with DFM, thermal simulation and EMC remediation in the same loop.

Enclosure and thermals treated as camera design

Sealing, anti-fog and sensor heat are validated together with image noise, because they are one problem in the field.

Clear boundary with the product brand

If you want an off-the-shelf unit rather than custom design, that belongs to MeToclaw. We will say so rather than sell you the wrong thing.

Related

Related Resources