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.
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.
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.
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.
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.