To consistently label two specific individuals (Person A and B) in a real-time CPR application using YOLOv11, follow this structured approach:

Solution Overview

1. Integrate a Tracking Algorithm (e.g., DeepSORT/SORT) to maintain persistent IDs across frames.
2. Use Heuristics (e.g., spatial position, pose keypoints) to differentiate between Person A and B.
3. Handle Edge Cases (occlusions, new people entering the frame) using tracking longevity and fallback logic.

Step-by-Step Implementation

1. Detection + Tracking Pipeline

• YOLOv11: Detect poses and bounding boxes for all people in the frame.
• DeepSORT/SORT: Assign unique IDs to detected people and track them across frames. Use:
  • Motion (Kalman Filter) for short-term predictions.
  • Appearance features (if using DeepSORT) to re-identify after occlusions.

2. Track ID Longevity

• Maintain a dictionary to track each ID’s "age" (number of consecutive frames it has appeared).

• Reset age to 0 if an ID is missing in a frame.

• Example data structure:

  tracked_persons = {
      id: {
          "age": int,           # Frames since first detection
          "last_bbox": list,    # [x1, y1, x2, y2]
          "last_keypoints": list # Pose keypoints
      }
  }
  
  

3. Select Primary Candidates (A and B)

• In each frame, select the two oldest IDs (highest age values) as candidates for A/B.
• If IDs are lost temporarily, retain their last known position for up to N frames (e.g., 5 frames).

4. Differentiate Person A vs. B

• Spatial Heuristic:

  • Sort candidates by horizontal position (e.g., leftmost = Person A).

  candidates = sorted(candidates, key=lambda x: x["last_bbox"][0])  # Sort by x1 of bounding box
  person_a = candidates[0]
  person_b = candidates[1]
  
  

• Pose Heuristic (CPR-specific):

  • Person A (compressions) has lower wrist keypoints (e.g., near the chest).
  • Person B (rescue breaths) has higher wrist keypoints (near the head).