How Low-Power AI Chips and Alibaba Cloud Power a Million-Scale Smart Camera Platform

Edge Device: Ultra‑Low‑Power AI Vision SoC

Research Microelectronics (研极微) provides a custom AI vision SoC that integrates a 4 TOPS NPU. The chip consumes about 60 % less power than conventional designs, enabling operation from solar panels or a few batteries while maintaining detection latency under 100 ms for categories such as humans, vehicles, pets, and flames.

Dual‑Path Data Upload

Video segments are uploaded through the SDK in a streaming fashion, while AI event notifications are sent via a separate MQTT channel. The two paths are independent and non‑blocking, allowing continuous video ingestion and timely event reporting.

Cloud Storage Backbone: OSS + Tablestore

Object Storage Service (OSS) stores raw video files. It offers 99.999999999 % (eleven 9s) data durability, multi‑region replication, and lifecycle policies that automatically tier hot data to lower‑cost storage classes and delete expired objects, reducing long‑term storage costs.

Tablestore holds structured metadata, device status, and AI‑generated embeddings. It provides high‑throughput, low‑latency (millisecond‑level) reads and writes, multi‑dimensional indexing, and vector search for embedding‑based retrieval. The service is serverless: partitions split and capacity expands automatically during traffic spikes, eliminating manual sharding.

Encryption and Privacy

All video data passes through an end‑to‑end encrypted pipeline (capture → transfer → store → playback). Each device receives an isolated storage namespace, preventing data leakage if a device or local storage is compromised.

Intelligent Event Detection and Cloud AI Summarisation

On‑device AI detects anomalies in real time and publishes alerts. Example scenarios include:

Elderly fall detection in residential settings.

Perimeter monitoring for outdoor areas.

Warehouse safety, prioritising fire (flame) detection and human presence.

Cloud AI further processes video to generate textual summaries (e.g., "3 PM – courier appears at the door") and supports natural‑language search of video segments, enabling keyword‑based retrieval such as "delivery person" or "red‑clothed woman".

Scaling to Tens of Millions of Devices

The serverless nature of Alibaba Cloud OSS and Tablestore accommodates growth from millions to tens of millions of edge devices without pre‑provisioned capacity. Automatic partitioning and elastic scaling support edge‑to‑cloud workloads, multimodal data fusion, and cross‑region deployments.