Case Overview
This project is aimed at building local face detection capabilities on smart glasses. The focus is not on identity recognition, but on the device side to complete face target detection in the direction of front face, left face, right face, etc., and output the detection frame, confidence, five key points, posture angle, direction classification and trackId within the same detection session to the upper-layer Android application. The project delivery form includes Android Face SDK, installable Demo APK, source code project, model file, access guide, error troubleshooting document and delivery status description. Public cases do not disclose customer names, on-site samples and internal evaluation data; the information on the page is used to explain the technical route, delivery content and project boundaries.
Project background and requirements
Smart glasses, mobile inspection terminals and edge vision devices often need to run vision algorithms under conditions where the network is unavailable, privacy constraints are strong, or the response path is short.Unlike cloud recognition, on-device offline detection requires a balance between device computing power, memory, camera frame format, Android life cycle and power consumption constraints. The core requirement of this project is to enable Android applications to access offline face detection capabilities on RK3588 smart glasses, support real-time camera footage and local image input, and output data structures that can be directly used for preview overlay, counting, direction judgment and subsequent business logic.
The requirements boundaries were clarified at the early stage of the project: the SDK provides face frames, confidence, five key points, yaw/pitch/roll posture angles, front face and left and right side face direction classification, synchronous and asynchronous detection interfaces, Camera2 real-time frames and local image detection examples; the SDK does not provide person registration, feature extraction, 1:1 or 1:N identity comparison, name or document recognition, cloud API, camera driver, USB Host or UVC Collection driver. Such a boundary helps customers distinguish between "detection capability delivery" and "identity recognition system delivery", and also facilitates subsequent phased expansion.
Technical route
The algorithm side adopts the offline CPU detection route, encapsulates the face detection capability based on the OpenCV Java API and YuNet model, and splits it into two AAR modules, Core and Face, in the Android project. Core AAR is responsible for general operation, status and basic tool capabilities, and Face AAR provides face detection interface, parameter configuration and result structure. The two AARs need to use the same SDK version and be connected with the Android dependency of OpenCV 4.10.0.The target platform is RK3588 Android device, the first-stage verification ABI is arm64-v8a, the minimum Android API is 24, compileSdk and targetSdk are 30.
The input layer supports YUV_420_888, NV21 and RGBA_8888. The real-time preview scene uses Camera2 to obtain 1280×720 YUV_420_888 frames. Before closing Image, the plane data is copied to its own DirectBuffer and then submitted to the SDK. For local image scenes, JPG or PNG can be converted to RGBA DirectBuffer for static detection. Preview mirroring, screen scaling and detection frame overlay belong to UI display layer mapping. The SDK returns the detection results based on the original input image coordinates. When drawing on the front end, it needs to be converted according to the actual preview area and whether to mirror.
Key points of Android SDK access
The SDK access process is based on the principle of "run the demo first, and then integrate the AAR". Customers can first install the Debug Demo APK, authorize camera permissions, verify the basic output of front face, left face, right face and multi-target scenes, and then incorporate Core AAR, Face AAR, OpenCV dependencies and model files into their own Android projects. Perform model installation and SDK creation when the application is initialized to avoid repeating install or create in each frame;It is recommended to use an asynchronous interface for real-time detection. A single SDK instance only retains one asynchronous task at the same time. If the previous frame task has not ended, it will return to the BUSY state, and the application side will drop frames or throttle according to business needs.
The Demo project adopts a detection submission interval of about 100 ms, which is approximately equivalent to a demonstration strategy of 10 times/second; this is not a fixed upper limit of SDK, and the actual detection frequency needs to be confirmed based on the RK3588 device load, input resolution, number of threads, screen complexity, and power consumption strategy. Detection configuration includes parameters such as the maximum number of faces, minimum face pixel size, confidence threshold, NMS threshold, whether to output poses and key points, whether to enable in-session tracking, etc. For side face detection, extremely close side faces, insufficient visible facial features, gestures beyond the sample distribution, motion blur, screen reflection, low illumination or rotation angle processing errors will affect detection stability and need to be verified based on samples and on-site conditions.
Delivery content and verification results
This delivery package is named EyeAlgo-Face-SDK-0.4.5-fix5-CUSTOMER_TEST-NOT_PRODUCTION-20260729, which is used for customer functional testing and Android project access evaluation, and is not used as a production signature version.The delivery content includes eye-algo-core-0.4.5-release.aar, eye-algo-face-0.4.5-release.aar, Eye-Algo-Face-Combined-Demo-0.4.5-fix5-debug-NOT_PRODUCTION.apk, FaceSdkProject, DemoProject, access guide, FAQ and error troubleshooting, delivery status description, third-party license and verification script.
| Verification items | Results | Description |
|---|---|---|
| Delivery package integrity | Passed | 138 files in the package match SHA256SUMS.txt, no additional unlisted files. |
| SDK source code construction | Succeeded in | JDK 11, Gradle 7.5.1, AGP 7.4.2 environment. |
| Unit test | Passed | Core 6/6, Face 1/1, Demo 13/13 passed. |
| Lint errors | Passed | 0 errors, 19 warnings, warnings need to be evaluated according to customer engineering specifications. |
| Production signature and mass production acceptance | Not included | Production signature, customer frozen sample accuracy and long-term stability acceptance need to be continuously confirmed in the customer environment. |
Reusable value
This case is suitable for evaluating offline face detection SDK delivery on smart glasses, mobile inspection terminals, edge computing boxes, Android industrial handheld terminals and other devices. Reusable content includes end-side model encapsulation, AAR modular delivery, Camera2 frame processing, real-time preview overlay, direction classification result output, static image detection, error code design, access documentation and verification checklist. For projects that need to complete visual perception locally but do not want to upload images to the cloud, this pattern can be used as a reference path for solution validation, prototype verification, and customer SDK delivery.
It should be noted that the public cases only present the verified project delivery facts, and do not write the unfinished customer official accuracy rate, 8-hour stability, production signature or mass production conclusion as achieved indicators. Formal projects still need to be independently verified based on customer frozen samples, real lens modules, lighting ranges, wearing angles, device temperature rise, application thread models and acceptance protocols.
FAQ
RK3588 smart glasses offline face detection SDK What is included in the delivery?
Contains Core AAR, Face AAR, Debug Demo APK, source code projects, model files, Android access guide, FAQs, delivery status instructions, license information and verification scripts, suitable for customer functional testing and engineering access evaluation.
Does the SDK include identity recognition function?
Not included. The scope of delivery is face detection and orientation classification, outputting face frames, confidence levels, key points, pose angles, and intra-session trackIds. Registration, feature libraries, 1:1 or 1:N identity comparison, name or document recognition are not provided.
What else needs to be verified before official launch?
It is necessary to confirm the detection accuracy, long-term stability, production signature, power consumption, temperature rise, camera drive, performance under different lighting and wearing postures on the customer's real equipment and samples, and form an acceptance record recognized by both parties.