Add speak detection model

[WenjinFu]

This pull request introduces several major improvements to the face recognition and tracking system, focusing on performance, robustness, and new features. The most significant update is the addition of a threaded camera capture mechanism to avoid frame drops and increase FPS. It also adds a vision-only "who is speaking" scorer (VVAD) and a more accurate head pose estimation module, along with API changes to leverage these features.

Camera capture performance and reliability:

• camera_reader.py: Refactored to use a background thread that continuously grabs the latest frame, ensuring read_frame() is non-blocking and always returns the freshest frame. This prevents frame drops and improves FPS by decoupling capture from processing. The threaded approach can be toggled via a parameter. [1] [2] [3] [4] [5] [6]

Vision-only speaking detection (VVAD):

• face_tracker.py, http_api.py: Introduced a vvad attribute that allows the system to track and score which detected face is likely speaking, based only on visual cues. The face tracker now feeds mouth/chin crops to the VVAD model, and the HTTP API can bias enrollment toward the visually detected speaker. [1] [2] [3] [4] [5] [6] [7]

Improved head pose and frontality scoring:

• head_pose.py: Added a new module for full head pose estimation (yaw, pitch, roll) using only five keypoints and optional chin landmark. This yields a more accurate "frontality" score, penalizing faces that look up or down, improving enrollment quality.

API and selection logic enhancements:

• http_api.py: Added a /speaking endpoint and logic to bias the selection of the enrollment face toward the one visually detected as speaking, ensuring the correct person is enrolled during multi-person interactions. [1] [2] [3]

These changes collectively make the system more robust, responsive, and accurate in real-time face recognition and speaker identification scenarios.