What is the compute of AIBox07?

The processor is AX650N. The CPU is an octa-core Cortex-A55 at 1.7GHz. The NPU is 18.0 TOPs INT8. Standard configuration is 8GB LPDDR4 memory and 64GB storage.

How many video channels can AIBox07 analyze at the same time?

Video decoding supports 32 channels of 1080p at 30fps, and the maximum decoding capability is 8K 60fps. Different cameras can be assigned different algorithm analysis tasks.

Which algorithms and video protocols does AIBox07 support?

It supports YOLO v5, YOLO v8, and multiple pretrained computer-vision models. Video ingest protocols include RTSP, WebRTC, RTMP, HTTP-FLV, Websocket-FLV, and HLS.

What are the dimensions and main interfaces of AIBox07?

The chassis measures 214mm × 120mm × 54mm. Interfaces include Gigabit WAN and LAN, HDMI 2.0, 1 USB 2.0, 1 USB 3.0, and a TF card, with an optional 4G/5G module.

AIBox07 Edge Video Analysis Inference Box

AI Box / March 2025

Model: AIBox-07 · Edge Inference Intelligent Computing Platform

The AI edge computing server box (AIBox-07) is based on a high-performance AI chip, with an octa-core ARM Cortex-A55 CPU and multiple integrated AI algorithms. Compute is up to 18T (INT8). It supports multiple deep-learning frameworks and can deploy AI to the edge, plug and play.

18T INT8
32 channels of 1080p
AX650N
YOLO v5 / v8

Front of the AIBox07 edge video analysis inference box, showing network ports, audio jacks, terminals, and HDMI

Features

Supports multiple pretrained AI computer-vision algorithm models.
Supports ingest over RTSP, WebRTC, RTMP, HTTP-FLV, Websocket-FLV, HLS, and similar protocols.
Algorithm models and the application platform in one, for integrated edge deployment.
Supports remote upgrade when the device is online.
Open interfaces, supporting secondary development and integration.
Supports algorithm models for the YOLO v5 and YOLO v8 deep-learning frameworks.
Supports multi-channel video stream aggregation (NVR).
Different cameras can be assigned different algorithm analysis tasks.
Supports local video storage (an SD card is required) and video playback.

Specifications

Product	AI edge computing inference box AIBox 07
System	ARM Linux (Ubuntu 22.04 or Debian 10)
CPU	High-performance embedded microprocessor AX650N, octa-core Cortex-A55 1.7GHz
NPU	18.0 TOPs @ INT8
Storage	Standard 8GB LPDDR4, 64GB storage
Video encoding	H.264 HP/MP/BP Level v5.1 encoding H.265 MP Level v5.0 encoding H.264/H.265 support I/P frames Maximum encoding capability: 8K @30fps H.264/H.265 multi-stream real-time encoding: 7680×4320 @30fps + 1080p @30fps JPEG snapshot: 3840×2160 @200fps CBR/VBR/AVBR/CVBR/FIXQ, P/QPMAP bitrate control Supports encoding of 8 regions of interest (ROI) Output bitrate up to 200Mbps
Video decoding	H.264 HP/MP/BP Level v5.2 decoding H.265 MP Level v5.1 decoding Maximum decoding capability: 8K @60fps Supports decoding of 32 channels of 1080p @30fps JPEG decoding: 3840×2160 @120fps Maximum decoding bitrate 300Mbps

Hardware Interfaces

Interface	Specification	Notes
LINE IN	Pink 3.5mm audio jack	Audio input
LINE OUT	Light green 3.5mm audio jack	Audio output
ALARM & RS485 & RS232	20pin 3.96mm Phoenix terminal	Wiring diagram is on the appearance drawing
WAN	RJ45	Supports 10/100/1000Mbps
LAN	RJ45	Supports 10/100/1000Mbps
HDMI	HDMI Type-A socket	Supports HDMI 2.0
4G/5G	M.2 interface	Supports a 4G/5G module, with a SIM card inserted
Storage	TF card	One TF card interface
Storage	M.2 NVMe interface	One (internal)
Storage	SATA interface	2 channels (internal)
USB	Two USB interfaces	One USB 2.0, one USB 3.0
DC12V	5.5mm / 2.0mm power jack
RESET	Round hole

Product Photos

AIBox07 front panel, with dual network ports, audio jacks, terminal block, HDMI, and power jack — Front-panel interfaces

AIBox07 rear panel, with TF card, M.2, and two USB ports — Dimensions: 214mm × 120mm × 54mm

Inference Capability

Model deployment and inference are packaged as capabilities for the intelligent-computing layer, with inference delivered as a platform and a service.

Deep learning flowchart of the training stage and the inference stage

Model deployment: model extraction, model compression, and model acceleration.

Model as a service: the model is deployed at the edge and integrated with the intelligent-computing application layer. The user sends image and video-stream requests from the intelligent-computing application layer; the model layer runs inference and returns the result.

Each video stream can run 3 to 5 models, and different video streams can be assigned different algorithms. 18T (INT8) compute supports 32 video streams.

Pretrained algorithm models

Multiple AI algorithms, with customization for the application scenario.

People Crowd gathering detection, illegal intrusion detection, person climbing, absence-from-post detection, monitoring-room absence detection, elevator entrapment detection, entry and exit detection, sleeping-on-duty recognition, smoking recognition, and other behavior analysis.

Vehicles Muck truck recognition, e-bike entering an elevator recognition, vehicle wrong-way recognition, non-motor vehicle illegal parking detection, vehicle illegal parking recognition, jaywalking recognition, heavy machinery recognition, pedestrian and vehicle separation recognition.

Objects Exposed earth pile detection, work uniform recognition, reflective vest recognition, face mask recognition, e-bike helmet recognition, insulating glove recognition, safety helmet recognition, fire extinguisher placement detection, protective clothing recognition, cat and dog recognition.

Environment Road litter recognition, flame recognition, flame detection, overflowing garbage recognition, corridor clutter recognition, open-door recognition for equipment rooms / power rooms / unit entrances, street-vending recognition, smoke recognition, smoke detection.

Intelligent Computing Platform

Video preview

Live view for channels 1–9.

Intelligent computing platform live preview for channels 1 to 9

Camera onboarding

Use an ONVIF management tool to search for cameras on the LAN and copy the video stream address. Then, in the intelligent computing platform, enter the device name and select the group name, channel number, protocol type, and video stream address to add the camera.

Intelligent analysis

Algorithm management provides multiple pretrained algorithm models.

Task management

Select different algorithms for different cameras and configure different tasks. Detection regions can be drawn and thresholds can be set.

Algorithm task configuration screen, where a detection region can be drawn and a threshold set

Alert records

Shows alarm images detected by algorithm analysis.

Video storage / playback

Management platform

The cloud management platform remotely manages, maintains, and upgrades AI edge computing boxes, including algorithm management, algorithm update delivery, and remote maintenance.

Management platform algorithm management screen

Management platform remote maintenance screen

Algorithm Customization

Custom development process for deep-learning AI vision algorithms:

1. Sample and annotate on site

Annotate images or videos sampled on site.

2. Build a dataset and train

After the dataset is obtained, design the model and start training.

3. Convert the deployment model

After the model is generated, convert it to an RKNN model for the RKNPU chip.

Algorithm models with distinct target features can be trained within 1 week. Behavior analysis takes about 3 to 4 weeks, with no fewer than 10000 samples.

Application Scenarios

Smart community / park

Algorithms and product applications: parking in a fire lane, area intrusion recognition, fall recognition, smoke and open flame recognition, boundary crossing detection, street vending, people counting, and elevator e-bike detection. They can be used for identity verification of people and vehicles at entry and exit, illegal parking, environmental sanitation, recognition of uncivil behavior, and fire-hazard detection.

Fire-lane parking detection — Parking in a fire lane

Street-vending recognition — Street vending

Villas, courtyards, and unattended sites

Algorithms and product applications: fire and smoke recognition, hazardous area recognition, loitering recognition, person detection, person lingering, illegal fishing / stranger intrusion, fall recognition, and license plate recognition. They can be used in villas, courtyards, fish ponds, farms, planting and breeding bases, and unattended stores.

Illegal fishing and stranger intrusion — Illegal fishing / stranger intrusion

Food-service stores

Algorithms and product applications: customer-flow counting, VIP customer recognition, smoking recognition, phone-call recognition, rodent recognition, chef uniform and hat detection, unattended open flame, staff procedure compliance, smoke and open flame recognition, and overflowing trash bins. They can be used in schools, food service, hospitals, and similar sites.

Smart campus

Student behavior analysis: jumping detection, running recognition, fall recognition, climbing recognition, fighting recognition, fire and smoke detection, smoking recognition, mobile phone recognition, tripwire detection, and area detection.

AI Video Analysis Edge Inference Box