BoxAI — edge AI for fixed cameras for your business

Three-layer architecture

• Fixed, PTZ or 360-degree IP cameras capture relevant locations.
• NVIDIA Jetson AI Boxes process video at the edge and publish events through MQTT/TLS or webhooks.
• A central VMS presents cameras, maps and alerts; drone feeds can be integrated to verify incidents.

Hardware and edge processing

Select the configuration based on camera count, algorithm mix and site survey findings.

• Jetson Orin NX 16GB Super is the primary box option; Orin Nano Super may suit smaller clusters.
• DeepStream handles the RTSP, video decoding, TensorRT inference and object-tracking pipeline.
• About 16 channels of 1080p decoding per box is a design estimate. Actual AI throughput depends on algorithm workload and requires on-site benchmarking.

Algorithms selected for each use case

The catalog contains 21 algorithms across asset protection, fire safety, public order, access control, intrusion and investigation. Not every algorithm runs simultaneously on every configuration.

• Asset protection: excavation, camera tampering, left/removed objects, vandalism and scene changes.
• Safety and public order: fire/smoke, audio events, fights, policy violations and unusual crowds.
• Access control: people counting, vehicle tracking, Vietnamese plate recognition, virtual tripwires and face lists.
• Intrusion and investigation: night intrusion, animals, unauthorized drones, PTZ tracking and weapons.
• Three platform capabilities: analytics dashboard, on-site deterrence and smart video search.

Match channel count to workload

Channel counts are design estimates, not a promise that every channel runs all 21 algorithms. Benchmark mAP and latency with the intended workload before confirming a configuration.

• Light workloads such as detecting and tracking people, vehicles or animals: up to an estimated 16 channels per box.
• Medium workloads with fire/smoke, tampering, tripwire or plate recognition: around 8–12 channels per box.
• Heavy workloads with pose, actions, face recognition or several combined algorithms: around 2–6 channels per box.

Packages by sector

• Heritage sites: prioritize excavation, night intrusion, fire/smoke, animals and artifact-store protection.
• Forestry and rangers: monitor fire, intrusion and patrol routes based on risk zones.
• Industrial sites and warehouses: focus on intrusion, fire, access and operational safety.

Deployment at Bao Ha Temple

NTA has delivered a deployment at Bao Ha Temple. This page shares the project qualitatively; camera counts, specific equipment and dates will be added only after confirmation.