Processing Pipeline

The VisionLog - AI Attendance module follows a dual-stage processing pipeline using YOLOv8 for person detection and tracking, combined with InsightFace for face detection and recognition.

Pipeline Overview

Step-by-Step Details

Step 1: Face Enrollment

Create the face database using folder-based batch enrollment.

Folder Structure:

• One folder per person
• Folder name = person identifier
• Multiple images per folder (3-5 recommended)

Enrollment Process:

1. Scan enrollment folder for person subfolders
2. Load all images from each person's folder
3. Detect face in each image (SCRFD)
4. Extract 512-d embedding (ArcFace)
5. Average embeddings for robustness
6. Save to embeddings.pkl database

Step 2: Video Input

Capture video or image input for processing.

Step 3: Person Detection

Component: YOLOv8

Detect all people in the frame before face processing.

Detection Output:

Why Person Detection First?

• YOLOv8 is fast and robust for detecting people at various distances
• Enables persistent tracking across frames
• Provides stable IDs even when faces aren't visible

Step 4: Tracking

Component: ByteTrack

Assign persistent track IDs to each detected person.

Tracking Features:

• Persistent IDs maintained across frames
• Motion prediction handles occlusions
• Re-identification when person re-enters frame
• Stable IDs even during partial occlusions

Step 5: Face Detection

Component: InsightFace SCRFD

Detect all faces in the full frame (not cropped to person boxes).

Detection Output:

Quality Filters:

• Minimum face size: 50 pixels
• Detection confidence: >= 0.5

Step 6: Face Recognition

Component: InsightFace ArcFace

Extract 512-dimensional face embedding for each detected face.

Embedding Properties:

• Dimension: 512
• Normalized: Unit vector
• Model: ArcFace (buffalo_l)

Step 7: IoU Matching

Component: Intersection over Union

Match detected faces to tracked persons based on spatial overlap.

Matching Process:

1. For each tracked person bounding box
2. Calculate IoU with each detected face bounding box
3. Assign face with highest IoU (if > threshold)
4. Link face identity to person track ID

Step 8: Identity Assignment & Attendance

Compare query embedding against enrolled database and log attendance.

Matching Process:

1. Compare query embedding against all enrolled embeddings
2. Find the highest similarity score
3. Accept match if score >= threshold

Threshold Guidelines:

Identity Caching:

• Last known identity stored per track ID
• Survives temporary face occlusions
• Only updates on confirmed (non-Unknown) matches

Attendance Logging:

• Frame number
• Timestamp
• Track ID
• Person name
• Confidence score
• Bounding box coordinates

Configuration

All settings are centralized for easy customization:

Data Flow Summary