Case 07 · Edge AI · Vision + Thermal

Small-Space People Counting
Vision-AI & Thermal-Imaging Reference Designs

From 60-mm ceiling-embedded AI sensors to wide-area thermal fire-safety detectors — three products, one playbook: edge NPU, no cloud, no stored video. Built on Rockchip RV1106B / RV1126B and reviewed against Bestom SoM and SDK.

Rockchip RV1106B RV1126B optional Edge NPU 1 TOPS <200 ms latency 98% accuracy Privacy-by-design
Why this matters

Use Cases

Existing hotels and office buildings are shifting energy and headcount efficiency from "card swipe / turnstile" to "contactless people counting". Three products map to three typical conditions.

Hotel guest room

Occupied / unoccupied triggers HVAC and lighting linkage, cutting energy 20–35%.

Small meeting room

Live occupancy display, linking to booking system and lighting control.

Small office / shared desk

Desk-occupancy statistics as a data source for work hours and scheduling.

Dorm / ward

Nighttime gathering / bed-exit detection, links to call system.

Multi-person ward visiting

Limits visitor count, alerts on anomalous gathering.

Battery warehouse / distribution room

Fire-prevention + people counting in one unit, IP65 housing optional.

Three products, one silicon family

Product Family

ModelApproachCore controllerNPUKey scenarioKey metrics
ZD101 vision-AI sensor RGB vision + head-shoulder detection ARM Cortex-A7 single-core 1 TOPS Small-space in/out & dwell Accuracy ≥98%, latency <200 ms, 4 persons/frame
RV1106B embedded solution RGB vision + head-shoulder detection Rockchip RV1106B 0.5–1 TOPS OEM/ODM custom finished unit Linux, WiFi+BLE, OEM/ODM ready
Thermal fire-safety detector Infrared thermal (128×96) Standalone MCU / swappable RV1126B Local inference Fire-prevention + privacy people counting 20–150 ℃ sensing, 40–60 m², RJ45
Product 01

ZD101 Vision-AI People-Counting Sensor

The housing needs only a 60-mm hole, ideal for hotel guest rooms, small meeting rooms and small offices. Local NPU inference outputs only count results — no video stored.

Key technical features
Vision analysisCamera builds a depth map; head-shoulder counting of in/out traffic
Data-fusion algorithmMulti-source weighting, accuracy up to 98%
Edge-compute architectureNPU chip, local latency <200 ms
Dynamic trackingSupports real-time trace of 4 objects/s
Privacy protectionAnonymized; no video stored, only statistics output
Key hardware parameters
ProcessorSingle-core ARM Cortex-A7
NPU performance1 TOPS
Memory / flash128 MB RAM · 64 MB FLASH
Operating systemLinux
Resolution200K px
Recognition speed6–7 fps
Concurrency4 persons/frame
MountingEmbedded, 60 mm hole
Mount height2.4 – 4.0 m
Field of viewH 79.1° · V 43.3° · D 94.9°
Power supplyDC 12 V
Output interfaceRS485 · UART · USB · Ethernet (scalable)
WirelessWiFi + Bluetooth
Power1.5 W
Operating temp−10 ~ +65 ℃
MaterialPlastic + aluminium alloy
Source: Zhendao Smart Technology ZD101 product manual V1.1 · debugging: remote / hotspot / debug APP (contact original manufacturer).
Product 02 · Reference Design

RV1106B Embedded Solution — Technical Architecture

A directly mass-producible embedded vision-AI terminal. This solution was baselined by the Bestom team together with the open-source community, and can quickly connect to customer protocols. Only technical composition and key components are listed below — no commercial quote.

ModuleKey componentsTechnical role
Main-control boardRockchip RV1106BSingle-core Cortex-A7 + 0.5–1 TOPS NPU, local vision inference
WirelessATBM6461 (WiFi/BLE)2.4G WiFi + BLE, connects cloud / APP
StorageW25N512VEIG (64 MB NAND)System image + model + config storage
ImagingCMOS sensor + IR-cut wide lens200K px imaging, low-light / black-light support
Interface boardUSB / RS485 / UART / SPI; DC 12 VExternal expansion & building/BA bus
IndicatorTri-color LED + IR emit/receiveStatus + near-field sensing
BluetoothExternal BLE5.0 (BCB02-7A)Near-field provisioning / remote
StructureAluminium lamp housing + partsHeatsink + −10~+65 ℃ wide-temp fit

Solution highlights

  • Rockchip RV1106B single-core A7 + 0.5–1 TOPS NPU, local inference
  • Linux OS, easy for secondary development & OTA
  • WiFi + BLE5.0 dual wireless, with RS485/UART/SPI to link BMS, elevator control and lighting
  • DC 12 V power, ~1.5 W level, can share power with fire / BA bus
  • Heatsink aluminium housing, −10 ~ +65 ℃ operating temp, fits distribution / server rooms

Why this chip

RV1106B is Rockchip's high-value SoC for entry-level vision AI: built-in ISP + NPU + light Linux, running 6–7 fps head-shoulder detection on a 200K-px camera. Bestom has already mass-produced vision-AI fire-prevention / smart-meeting terminals on RV1106B, so the solution transfers directly to people counting.

Product 03

Thermal Fire-Safety Detector (thermal-imaging route)

For privacy-sensitive sites where cameras cannot be installed: battery warehouses, distribution rooms, dorms, bathroom doors. Using infrared thermal imaging, it both alarms on temperature and roughlyrecognizes region occupancy while avoiding privacy risk.

Thermal fire-safety detector
Image source: thermal-detector promo page (product direction; final delivered form can be OEM).
Product positioning
Anti-bullying managementRegion-gathering alert, links audio alarm / light
People-gathering alarmGathering detection with settable threshold
High-temp appliance detection−20–150 ℃ sensing, for battery warehouse / charging
Fire-risk preventionEarly Li-battery thermal-runaway detection
High-temp source alarmCustom temperature alarm condition
Hardware parameters
Imaging techSingle-spectrum infrared thermal
Resolution128 × 96 (multiple optional)
CoverageSingle device covers 40 – 60 m²
Temp range−20 ~ +150 ℃
Output interfaceRJ45 wired (wireless optional)
Energy monitoringLinks HVAC/lighting, detects hidden appliances
Privacy protectionInfrared blurs privacy features; no video stored, only frame to ops
Why Bestom

Why BesTom

The three products use different silicon, but the step to mass production and protocol integration is the same thing — which is exactly BesTom's IDH core business.

One-stop Rockchip selection

From RV1103 / RV1106B (entry-level vision AI) to RV1126B (multi-camera), RK3572 (high compute), RK1828 (AI NPU co-processor) — no multi-supplier sourcing per project.

BSP / SDK / model acceleration

Linux BSP, ISP tuning, YOLO/Head-Shoulder model quantization & RKNN conversion on RV1106B / RV1126B already done — cuts 8–12 weeks of pre-study.

Small-space people-counting reference design

Bestom's ZD101derived solution already in small-batch shipment for hotel guest rooms, directly linking BACnet / Modbus / MQTT building systems.

Protocol integration & certification

Integrated BACnet, KNX, Modbus, MQTT, HTTP/Webhook. Optional CE / FCC / RoHS / EN50155 certification partners.

Mature mass-production chain

Shenzhen/Dongguan lines support customers from 50-unit pilot run to 30K/month. Already mass-producing massage-chair electronics, consumer AI fire-prevention, commercial AI terminals.

Manufacturability & DFM

Shenzhen/Dongguan lines ramp from pilot to mass production, with DFM report, thermal simulation & EMC remediation.

Related platforms on this site

Related SoM / Docs

Rockchip full-line SoM

RV1106B / RV1126B / RK3572 / RK1828 core boards & carrier boards, with open schematics.

Building / hotel / edge-AI solutions

Already mass-produced edge-AI vision & building-protocol integration solutions.

BSP / SDK / Datasheet

RV1106B BSP, Rockchip SDK, NPU deployment notes.

Inquire

Quote, NRE, pilot run & certificationcadence — bring your scenario & target BOM.