Home/Cases
AI Algorithm Delivery Case

RK3588 Smart Glasses Offline Frontal/Profile Face Detection and Android SDK Delivery

This case covers local face detection for smart glasses, including frontal/profile detection packaging, Android AAR SDK delivery, Camera2 real-time preview integration, demo project, customer test guide and checksum materials. It describes delivery facts only and does not claim identity recognition, face library or person search capability.

RK3588 Android edge runtimeOffline face boxes and confidenceAAR SDK and demo projectClear customer test boundary

Case Facts

Project type
Edge AI algorithm and Android SDK delivery for RK3588 smart glasses.
Hardware and OS
RK3588, Android 12, arm64-v8a, with a Camera2 external camera path.
Algorithm scope
Detection on real-time frames and local images, with face boxes, confidence, five landmarks, pose and per-session trackId.
Deliverables
Core/Face AARs, faceDebug demo project, integration notes, customer test guide, checksum files and verification records.
Boundary
No identity enrollment, feature database, 1:1 verification or 1:N person search API.
Verification boundary
Local build, package checks, Camera2 smoke tests and sample regression were completed. Formal accuracy depends on the customer frozen dataset and device conditions.

Project Background

The customer needed local face detection on smart glasses, returning frontal/profile face positions and algorithm results for the Android application layer. The core algorithm was required to run offline and reduce reliance on cloud recognition services or unstable site networks.

Edge real-time previewLocal image test pathReproducible SDK integration path

Technical Implementation

The SDK wraps an OpenCV 4.10.0 FaceDetectorYN/YuNet route as Android AARs and integrates Camera2 YUV_420_888 input, EXIF orientation, autofocus state, brightness and sharpness prompts. Async detection uses single-task backpressure to avoid queued work before the previous callback returns.

Camera2 external camera previewOwned DirectBuffer copy for Y/U/V planesBUSY(1602) backpressure handlingCoordinate mapping and preview overlay

Deliverables

The delivery covers algorithm SDKs, demo project, integration notes, customer test guide and verification materials. The AAR targets arm64-v8a. The demo supports real-time Camera2 detection, local JPG/PNG detection, standard mode and difficult-image enhancement mode.

Android AAR SDKfaceDebug demo projectGradle integration notesSHA256 checks and test records

Verification and Limits

The delivery stage completed RK3588 Android Camera2 smoke tests, async submit/callback consistency checks, resource release checks, and regression with synthetic frontal, synthetic 45-degree profile and no-face object samples. Research candidates and desktop latency were kept as engineering tuning data, not formal customer accuracy conclusions.

Desktop tests are not RK3588 formal performanceResearch candidates are not a customer acceptance setProduction APK signing requires the customer release key

Applicable Scenarios

This delivery pattern fits Android smart terminals, local AI boxes and edge devices that need face location detection for overlays, interaction triggers or downstream algorithms. Identity authentication, face libraries or security use cases require separate model, data, permission, compliance and acceptance evaluation.

Smart-glasses edge visionAndroid camera algorithm SDKOffline AI detection integrationEdge device vision assessment

Need RK3588 or Android edge AI assessment?

Submit the hardware platform, camera path, sample type and target features so the delivery scope and validation method can be confirmed first.

Request Assessment