{
  "schema_version": 1,
  "project": {
    "name": "fall-detection",
    "version": "1.0.0",
    "task": "XGBoost fall classification from 56-feature pose vectors (image → YOLO upstream)",
    "python": ">=3.10"
  },
  "paths": {
    "pose_model": "../yolo26x-pose.pt",
    "classifier_model": "models/xgboost_priority1_fall_model.pkl",
    "output_dir": "output"
  },
  "models": {
    "pose_model_type": "Ultralytics YOLO-Pose (upstream feature generation only, not classifier input)",
    "classifier_type": "XGBClassifier",
    "class_names": {
      "0": "Normal",
      "1": "Fall"
    },
    "feature_count": 56,
    "classifier_feature_count": 56,
    "csv_total_columns": 58,
    "column_breakdown": "51 pose features + 5 additional Priority 1 features + 1 split + 1 label = 58 columns; XGBoost uses only 51 + 5 = 56 (columns 1-56)",
    "row_construction": "row = norm_kpts (51) + [aspect_ratio] + p1_features (4) + [split, label]  → 51 + 5 + 1 + 1 = 58"
  },
  "thresholds": {
    "yolo_confidence": 0.25,
    "keypoint_confidence": 0.5,
    "fall_probability": 0.7,
    "_notes": "yolo_confidence is upstream YOLO detection threshold (not a classifier parameter); keypoint_confidence < threshold → NaN x/y in 56-vector (preserved confidence); fall_probability is XGBoost P(Fall) decision threshold"
  },
  "feature_schema": {
    "keypoint_count": 17,
    "values_per_keypoint": [
      "x_normalized",
      "y_normalized",
      "confidence"
    ],
    "keypoint_feature_count": 51,
    "bounding_box_feature_count": 1,
    "priority1_feature_count": 4,
    "additional_feature_count": 5,
    "additional_features": [
      "aspect_ratio",
      "nose_relative_y",
      "torso_angle",
      "norm_com_y",
      "head_hip_v_dist"
    ],
    "classifier_feature_count": 56,
    "classifier_columns": "1-56",
    "metadata_columns": [
      "split",
      "label"
    ],
    "metadata_count": 2,
    "metadata_indices": {
      "split": 57,
      "label": 58
    },
    "csv_total_columns": 58,
    "csv_total_breakdown": "51 pose (cols 1-51) + 5 additional (cols 52-56) = 56 classifier features; + 1 split (col 57, train/val/test) + 1 label (col 58, 0=Normal/1=Fall) = 58",
    "row_construction": "pose_extract.py: row = norm_kpts (51) + [aspect_ratio] + p1_features (4) + [split, label]  → 51 + 5 + 1 + 1 = 58; XGBoost uses only 51 + 5 = 56 (columns 1-56)",
    "column_indices": {
      "pose_features": "1-51",
      "additional_features": "52-56",
      "split": 57,
      "label": 58
    },
    "total_feature_count": 56,
    "low_confidence_coordinates": "NaN",
    "low_confidence_confidence_value": "preserved",
    "bounding_box_features": [
      "aspect_ratio"
    ],
    "priority1_features": [
      "nose_relative_y",
      "torso_angle",
      "norm_com_y",
      "head_hip_v_dist"
    ]
  },
  "input": {
    "pipeline_csv_input": {
      "type": "pandas.DataFrame",
      "consumer": "scripts/run.py",
      "description": "Upstream delivers a DataFrame with 51 normalized keypoint features + bounding box (55 columns total) per row; run.py extracts the 5 Priority 1 features (NOT input), builds a 56-column DataFrame (51+5), and passes that DataFrame (not a CSV) to XGBoost",
      "keypoint_columns": "x_0,y_0,conf_0,...,x_16,y_16,conf_16 (51 columns, x/y normalized to bbox; NaN allowed for occluded)",
      "keypoint_feature_count": 51,
      "bbox_columns": ["x1", "y1", "x2", "y2"],
      "bbox_feature_count": 4,
      "input_column_count": 55,
      "input_breakdown": "51 keypoint (x_i,y_i,conf_i i=0..16) + 4 bbox (x1,y1,x2,y2) = 55 total input to run.py",
      "bbox_units": "pixels",
      "computed_in_run_py": ["aspect_ratio", "nose_relative_y", "torso_angle", "norm_com_y", "head_hip_v_dist"],
      "extracted_count": 5,
      "model_input": "pandas.DataFrame, n rows × 56 columns (FEATURE_COLS = 51 input + 5 extracted)",
      "model_feature_count": 56,
      "output_columns": ["prediction", "fall_probability"],
      "min_valid_keypoints": 5,
      "cli_note": "CSV path on CLI is only upstream transport; it is loaded into a DataFrame before feature extraction"
    },
    "classifier_input": {
      "type": "feature_vector",
      "dimensions": 56,
      "dtype": "float",
      "missing_value": "NaN",
      "description": "Direct XGBoost input: 56-column pandas DataFrame (one row per person) = 51 pose features (17 keypoints × x_norm, y_norm, confidence) + 5 additional Priority 1 features; no image or CSV input at inference",
      "column_breakdown": "51 + 5 = 56 classifier features (cols 1-56); cols 57-58 (split, label) are metadata, not input",
      "very_important": "XGBoost does NOT use split or label as input features; label is supervision only"
    },
    "upstream_image_input": {
      "_note": "Images are processed upstream by YOLO26x-Pose to generate the 56-vector; not direct classifier input",
      "formats": [
        ".jpg",
        ".jpeg",
        ".png",
        ".bmp"
      ],
      "color_format": "BGR",
      "default_image": null
    }
  },
  "output": {
    "classifier_output": {
      "type": "binary_label",
      "labels": ["Normal", "Fall"],
      "label_ids": {"Normal": 0, "Fall": 1},
      "decision_rule": "P(Fall) >= fall_probability (0.7) → Fall, else Normal",
      "probabilities": ["P(Normal)", "P(Fall)"]
    },
    "save_annotated_default": false,
    "filename_prefix": "annotated_"
  }
}