How to personalize smart home with familiar face identification

Fast and accurate familiar face identification enhances smart homes by providing seamless convenience and personalized experiences tailored to each household member. This technology instantly recognizes familiar faces and strangers alike, enabling automatic adjustments to home settings based on individual preferences. By protecting biometric data locally and requiring minimal maintenance, Familiar Face Identification makes everyday interactions more intuitive and personalized.

embedUR combined edge AI with the STM32N6 MCU to tackle these challenges. The solution deploys a full facial recognition pipeline directly on-device, eliminating cloud dependency and the need for external hardware accelerators. Registration is modernized through a mobile companion app, enabling quick onboarding and seamless, secure transmission of facial data to the embedded smart lock. All matching and management run locally for instant access and reliable protection.

Application domains

• Personalized smart lighting and climate control
• Customized multimedia and entertainment settings
• Smart kitchen and appliance personalization
• Health and wellness monitoring and assistance
• Tailored notifications and reminders for household members

Traditional people identification systems have long relied on cloud-based recognition, RFID or keypad entry, and face recognition with external hardware accelerators. While these approaches provide foundational functionalities, each carries inherent limitations that affect user experience, security, and operational efficiency.

The limitations of legacy solutions inspired a new approach using lightweight facial recognition models running directly on STM32 microcontrollers. This innovation enables real-time, privacy-first biometric authentication without relying on cloud services or external accelerators. A mobile app companion modernizes user onboarding, providing intuitive face registration and secure transmission of biometric features to the embedded device.

Key benefits of the STM32 edge AI solution:

• Full on-device inference: eliminates cloud dependency, reducing privacy risks, latency, and operational costs
• Smooth user experience: mobile app simplifies the registration process with instant feedback
• Low power & compact: enables integration into battery-powered and space-constrained devices
• Fast response: achieves authentication within approximately 70 milliseconds

Real-time face recognition system that captures images from a camera, detects and validates faces, extracts embeddings, and either matches them to existing records or registers them as new entries. The pipeline emphasizes reliability through rigorous quality checks and efficient face-matching logic.

1. Image acquisition

Captures an RGB image from the camera, with a resolution of 224x224x3.

2. Face detection

A YOLOv8-based model identifies faces in the image, outputting bounding boxes and facial keypoints. Low-confidence detections are filtered out based on a predefined threshold, and Non-Maximum Suppression (NMS) is applied to remove overlaps and retain the most probable detections.

3. Quality assessment

Detected faces undergo multiple validation checks to ensure suitability for recognition:

• Keypoint confidence – Validates the reliability of detected facial landmarks.
• Nose position – Confirms the nose is within expected bounds, indicating frontal pose.
• Head tilt – Rejects faces with excessive vertical tilting.
• Distance checks – Ensures the face is neither too close nor too far from the camera, based on face size.

Only faces that pass all these criteria proceed to the next stage.

4. Face preprocessing

Valid faces are cropped, resized and aligned to correct orientation. An additional tilt-angle check is performed to discard excessively rotated faces. Preprocessed faces are then forwarded for recognition.

5. Feature extraction

The aligned face is passed through an ArcFace-based model to generate a 512-dimensional feature embedding that uniquely represents the individual.

6. Identity matching

The generated embedding is compared with a database of stored embeddings to determine identity:

Match found:

• If the similarity score exceeds the threshold and an identity is known, the name is displayed.
• If the identity is unknown, the user is prompted to assign a name.

No match (high-quality face):

• If no match is found but the face quality is sufficient, it is added to the database as a new, unnamed entry.

The MB1854B camera module, powered by the IMX335 CMOS image sensor, offers high-resolution imaging that captures sharp and detailed RGB images. The IMX335 is ideal for computer vision tasks, such as real-time face detection and recognition. The low-light performance & image clarity makes it reliable for capturing clear data in a variety of lighting conditions. The sensor produces RGB frames that serve as the input for face recognition systems.