TouchVision edge AI computers run deep-learning video analytics next to the cameras, so sites can detect fire, smoke, intrusion, missing PPE and unsafe behaviour in real time without streaming video to the cloud. The boxes reuse existing IP cameras and NVRs and push only events and alarms to your platform.
Platform
All current models are built on SOPHGO AI processors:
- BM1684: 8× Cortex-A53 at 2.3 GHz, 17.6 TOPS INT8, up to 32 channels of 1080p decoding and 16 concurrent analytics streams.
- BM1684X: 32 TOPS INT8 with FP32/FP16/BF16/INT8 mixed precision, used in the YZ-B033 for both video AI and LLMs.
- Efficiency: under 1 W per TOPS, with passive-cooling options on selected designs.
- Frameworks: TensorFlow, PyTorch, ONNX, Caffe, MXNet and PaddlePaddle; Docker, Python, OpenCV and FFmpeg hardware acceleration.
Choose a model
| Model | Key difference |
|---|---|
| YZ-B011 | 3× GbE, isolated CAN, RS-485, DIO, DC 12–36V wide input |
| YZ-B012 | 32-channel decode for camera-dense sites |
| YZ-B013 | Compact, ≤20 W, modular SATA/LTE/Wi-Fi/HDMI options |
| YZ-B015 | M.2 NVMe plus optional 2 TB storage |
| YZ-B016 | Direct 220V AC input, alarm inputs |
Ready-made algorithms
More than 50 algorithms are available, including fire and smoke, intrusion and line crossing, crowd gathering, helmets and work clothing, phone use and smoking, illegal parking, blocked fire lanes, e-bikes entering elevators, overflowing bins, sleeping or absent guards, falls and objects thrown from height. Scenario packs cover fuel stations, commercial kitchens, residential communities, battery charging sheds, construction sites, factories, warehouses, ports and mines. Custom models can be trained and deployed on request.
Integration
Industrial I/O (RS-485, isolated CAN, opto-isolated DI/DO) lets a box trigger sirens, barriers or a PLC directly. Results can be shown on a TouchVision security & fire-safety terminal or sent to your VMS over the network. Learn more in AI video analytics solutions.
Products in this category
- YZ-B011 16-Channel AI Video Analytics Box (BM1684, 17.6 TOPS): An industrial edge AI box built on the BM1684 (17.6 TOPS INT8) that analyses 16 HD camera streams at once, with 3× GbE, isolated CAN, RS-485, DIO and 12–36V DC input.
- YZ-B012 32-Channel AI Edge Server (BM1684): A BM1684 AI edge server (17.6 TOPS INT8) for camera-dense sites: decodes 32 channels of 1080p and runs real-time analytics with wide-voltage industrial I/O.
- YZ-B013 Compact AI Edge Computing Box (BM1684): A compact, ≤20 W edge AI box (BM1684, 17.6 TOPS) that analyses 16 streams and decodes 32× 1080p, with modular options for SATA, LTE, Wi-Fi and HDMI.
- YZ-B015 AI Edge Server with NVMe Storage (BM1684): A small-footprint BM1684 AI edge server with M.2 NVMe and up to 2 TB optional storage for keeping evidence clips alongside real-time analytics.
- YZ-B016 AC-Powered AI Video Analytics Box (BM1684): A BM1684 AI video analytics box that runs directly from 220V AC mains, with alarm inputs, dual RS-485 and HDMI for control-room installations.
Request a quote for standard or customized versions.
Frequently asked questions
What is an edge AI computer?
An edge AI computer is a compact industrial computer with a dedicated AI accelerator that runs neural-network inference on site, close to cameras or sensors, instead of sending data to a cloud server.
How many cameras can one TouchVision AI box handle?
BM1684-based models analyse up to 16 HD streams at the same time and decode up to 32 channels of 1080p video. Actual capacity depends on the algorithm mix.
Can I use my existing cameras?
Yes. The boxes take standard RTSP streams from existing IP cameras and NVRs, so no camera replacement is needed.
Can I run my own AI model?
Yes. Models from TensorFlow, PyTorch, ONNX, Caffe, MXNet and PaddlePaddle can be converted for the SOPHGO toolchain and deployed, with Docker and Python support.