Edge AI
AI that decides right on the device.
On-device Real-time Inference
Edge AI doesn't ship every byte to the server — it recognizes and decides right on the user's device or on-site equipment.
AX Lab researches how on-device detection, real-time inference, and lightweight models connect to smart-city, spatial data, and AR interfaces.
— Section 1
Three conditions are core.
01 On-Device Object Detection
Detect people, objects, signs, and hazards inside the device, using camera and sensor input. The starting point for field-grade AI that must keep judging even when the network is unstable.
02 Real-Time Edge Inference
Decisions must arrive without delay — the moment the user moves, the screen changes, or a vehicle or device shifts. For edge inference, response speed is usability and safety.
03 Quantization
Make models run on constrained hardware. The craft is balancing model size, accuracy, speed, and battery consumption.
— Section 2
Extend into AI structures that work on-site.
01 Urban Data Platform × Edge AI
Research that combines urban data, AR, and on-device AI to connect field-sensed information to service data. Possibility: Quickly judge traffic, facilities, safety, and environment data in the field; connect to city-operations dashboards as smart-city infrastructure.
02 AR HUD / Smart-Glass Recognition
An interface that places camera-based recognition results on the user's field of view. When recognition and display happen together, edge AI's value becomes clear.
03 On-device Content Review
Review basic quality, objects, and rule violations on the device at the moment the image or video is generated or captured. Possibility: Automate first-pass review before upload — for creative production, field capture, and product registration — cutting operational cost.
— Section 3
Edge AI tags are on-site execution conditions.
On-Device
Decisions inside the device, no server round-trip. On-device content review and field detection do the first-pass judgment locally — cutting network latency and the burden of sending personal data. Related projects: On-device Content Review, Field Object Detection
Real-Time
React to user behavior and on-site conditions instantly. AR HUD / Smart-Glass interfaces fall apart if guidance arrives late. Real-time inference is the core requirement to reflect location, heading, and objects immediately. Related projects: AR HUD / Smart-Glass, Real-time City Guidance
Quantization
Lightweight models that run on small devices. On mobiles, kiosks, and glass devices, model size and power efficiency matter. Quantization is a deployment technique that balances accuracy and speed. Related projects: On-device Inference Optimization, Edge Deployment Pipeline
— Timeline
Fast AI works close to the user.
Near · Zero-latency recognition
Interpret user scenes and behavior instantly, without a server round-trip.
Next · Field automation
Capture, inspect, guide, and review — automatically, on-site.
Future · Distributed AI infrastructure
Build service architectures where devices, spaces, and servers share judgment roles.