data_utils.py

import numpy as np

def normalize_landmarks(landmarks):
    """
    Converts raw landmarks to body-relative coordinates.
    Input: Flat array or list of (Pose + LH + RH) coordinates.
    Output: Normalized flat array.
    """
    # 1. Reshape to (N, 3) for easier math
    # Structure: 33 Pose + 21 LH + 21 RH = 75 points
    # Each point has x, y, z
    reshaped = np.array(landmarks).reshape(-1, 3)
    
    # 2. Extract Key Body Points (MediaPipe Pose Indices)
    # 11: Left Shoulder, 12: Right Shoulder
    pose_start = 0
    left_shoulder = reshaped[pose_start + 11]
    right_shoulder = reshaped[pose_start + 12]
    
    # 3. Calculate Center (Midpoint of shoulders)
    # This becomes our new (0,0,0)
    center = (left_shoulder + right_shoulder) / 2
    
    # 4. Calculate Scale (Shoulder width)
    # We use this to divide all coordinates so 'distance' doesn't matter
    shoulder_width = np.linalg.norm(left_shoulder - right_shoulder)
    
    # Avoid division by zero if shoulders are not found
    if shoulder_width < 0.01:
        shoulder_width = 1.0
        
    # 5. Normalize
    # (Point - Center) / Scale
    normalized = (reshaped - center) / shoulder_width
    
    # 6. Flatten back to 1D array
    return normalized.flatten()