column_descriptions.md

Kaggle column descriptions - copy/paste reference

For each group below, open every listed file's "Column descriptions" editor on the Kaggle dataset page and paste in the same column/description text.

Label files (start_time / end_time / action / is_fall)

Files:

• train/labels/hr_fall_detection_1.csv
• train/labels/hr_fall_detection_2.csv
• train/labels/hr_fall_detection_3.csv
• valid/labels/fall_detection_4.csv
• valid/labels/fall_detection_5.csv
• valid/labels/fall_detection_6.csv
• valid/labels/fall_detection_7.csv
• valid/labels/fall_detection_8.csv
• valid/labels/fall_detection_9.csv
• valid/labels/fall_detection_10.csv

Columns:

• start_time (integer): Start of the labelled segment, in seconds from the start of the video.
• end_time (integer): End of the labelled segment, in seconds from the start of the video.
• action (string): Activity/posture label for this segment (e.g. Stand, Sit, Lie). A hyphenated value (e.g. Stand-Sit) marks a transition between two states. Blank means no activity was assigned yet (e.g. before the subject enters frame).
• is_fall (boolean): Whether this segment is labelled as a fall event.

Static-pose feature files (un-predicted)

Files:

• train/pose-1st-iteration/features_output.csv
• train/pose/features_output.csv
• train/pose/model1_features_output.csv
• train/pose/model2_features_output.csv

Columns:

• image_name (string): Source static-pose image file name.
• is_upright (boolean): Whether the shoulder line is rotated enough relative to the hip line to indicate an upright torso (part of the standing-vs-lying rule).
• percent_upright (numeric): Confidence percentage backing the is_upright classification.
• stand_left (string): Left-leg standing classification from hip-knee-ankle bone-length geometry: standing, non_standing, squat, or uncertain_standing/uncertain_squat.
• stand_right (string): Right-leg equivalent of stand_left.
• percent_stand_left (numeric): Confidence percentage backing the stand_left classification.
• percent_stand_right (numeric): Confidence percentage backing the stand_right classification.
• sit_left (string): Left-leg sitting classification: sitting, non_sitting, or uncertain_sitting.
• sit_right (string): Right-leg equivalent of sit_left.
• percent_sit_left (numeric): Confidence percentage backing the sit_left classification.
• percent_sit_right (numeric): Confidence percentage backing the sit_right classification.
• lie_left (string): Left-side lying classification: lying or non_lying.
• lie_right (string): Right-side equivalent of lie_left.

Static-pose feature files (predicted)

Files:

• train/pose-1st-iteration/features_output_predicted.csv
• train/pose/features_output_predicted.csv
• train/pose/model1_features_output_predicted.csv
• train/pose/model2_features_output_predicted.csv

Columns:

• image_name (string): Source static-pose image file name.
• is_upright (boolean): Whether the shoulder line is rotated enough relative to the hip line to indicate an upright torso (part of the standing-vs-lying rule).
• percent_upright (numeric): Confidence percentage backing the is_upright classification.
• stand_left (string): Left-leg standing classification from hip-knee-ankle bone-length geometry: standing, non_standing, squat, or uncertain_standing/uncertain_squat.
• stand_right (string): Right-leg equivalent of stand_left.
• percent_stand_left (numeric): Confidence percentage backing the stand_left classification.
• percent_stand_right (numeric): Confidence percentage backing the stand_right classification.
• sit_left (string): Left-leg sitting classification: sitting, non_sitting, or uncertain_sitting.
• sit_right (string): Right-leg equivalent of sit_left.
• percent_sit_left (numeric): Confidence percentage backing the sit_left classification.
• percent_sit_right (numeric): Confidence percentage backing the sit_right classification.
• lie_left (string): Left-side lying classification: lying or non_lying.
• lie_right (string): Right-side equivalent of lie_left.
• label (string): Ground-truth static posture (stand, sit, lie, bend, squat, etc.), parsed from the image filename.
• predicted_label (string): The package's rule-based static-posture prediction for this image. Blank where MediaPipe failed to detect any landmarks (this happens on some lying/occluded poses).

K-means aspect-ratio cluster summary

Files:

• train/pose/grouped_kmeans_static_pose_boundingbox_data.csv

Columns:

• aspect_ratio (numeric): Minimum bounding-box aspect ratio within this K-means cluster. NOTE: this file is a pandas multi-index aggregation (groupby(...).agg(['min','max','mean','count'])) exported with duplicate-suffixed column names; row 0 of the data holds the literal sub-column labels ('min','max','mean','count') rather than a value - skip row 0 when loading, or treat it as a second header row.
• aspect_ratio.1 (numeric): Maximum bounding-box aspect ratio within this K-means cluster (see column 'aspect_ratio' note about row 0).
• aspect_ratio.2 (numeric): Mean bounding-box aspect ratio within this K-means cluster (see column 'aspect_ratio' note about row 0).
• aspect_ratio.3 (numeric): Number of images in this K-means cluster (see column 'aspect_ratio' note about row 0).

Overall bounding-box summary

Files:

• train/pose/grouped_static_pose_boundingbox_data.csv

Columns:

• aspect_ratio (numeric): Mean bounding-box aspect ratio across all 113 static pose images. NOTE: same pandas multi-index export artifact as grouped_kmeans_static_pose_boundingbox_data.csv - row 0 of the data holds the literal sub-column label ('mean') rather than a value.
• relative_width (numeric): Mean bounding-box width as a fraction of frame width, across all 113 images (see 'aspect_ratio' column note about row 0).
• relative_height (numeric): Mean bounding-box height as a fraction of frame height, across all 113 images (see 'aspect_ratio' column note about row 0).

Bounding-box geometry (focused keypoints)

Files:

• train/pose/static_pose_boundingbox_data.csv

Columns:

• image_name (string): Source static-pose image file name.
• frame_width (integer): Source image width, in pixels.
• frame_height (integer): Source image height, in pixels.
• aspect_ratio (numeric): Height/width ratio of the MediaPipe pose bounding box.
• relative_width (numeric): Bounding box width as a fraction of the frame width.
• relative_height (numeric): Bounding box height as a fraction of the frame height.

Bounding-box geometry (all 33 keypoints)

Files:

• train/pose/static_pose_boundingbox_data_all_keypoints.csv

Columns:

• image_name (string): Source static-pose image file name.
• frame_width (integer): Source image width, in pixels.
• frame_height (integer): Source image height, in pixels.
• aspect_ratio (numeric): Height/width ratio of the MediaPipe pose bounding box.
• relative_width (numeric): Bounding box width as a fraction of the frame width.
• relative_height (numeric): Bounding box height as a fraction of the frame height.

Bounding-box geometry (with prediction)

Files:

• train/pose/static_pose_boundingbox_data_bounding_box_size_predicted.csv

Columns:

• image_name (string): Source static-pose image file name.
• frame_width (integer): Source image width, in pixels.
• frame_height (integer): Source image height, in pixels.
• aspect_ratio (numeric): Height/width ratio of the MediaPipe pose bounding box.
• relative_width (numeric): Bounding box width as a fraction of the frame width.
• relative_height (numeric): Bounding box height as a fraction of the frame height.
• label (string): Ground-truth static posture, parsed from the image filename.
• predicted_label (string): Posture prediction from a bounding-box-aspect-ratio-only classifier - an alternative to the landmark-angle-based classifier, useful when MediaPipe landmark detection itself fails (e.g. on occluded or lying poses).

Per-frame video results

Files:

• train/results of static pose classifier on training videos/hr_fall_detection_static_pose_class_on_video_1_results.csv
• train/results of static pose classifier on training videos/hr_fall_detection_static_pose_class_on_video_2_results.csv
• train/results of static pose classifier on training videos/hr_fall_detection_static_pose_class_on_video_3_results.csv

Columns:

• file_name (integer): Frame index within the source video.
• max_value (numeric): Peak frame-differencing magnitude used to flag motion at this frame.
• process_image (boolean): Whether this frame was selected for pose estimation. Frames are motion-triggered rather than exhaustively sampled, so not every frame is processed.
• label (string): Ground-truth posture for this frame, derived from the manual activity-label CSV for the matching video and timestamp.
• prediction (string): The package's rule-based static-posture prediction for this frame. Blank where the frame was not processed or no pose was detected.
• image_name (string): Unused in this per-frame file (carried over from the shared feature-extraction function); always blank.
• is_upright (boolean): Whether the shoulder line is rotated enough relative to the hip line to indicate an upright torso (part of the standing-vs-lying rule).
• percent_upright (numeric): Confidence percentage backing the is_upright classification.
• stand_left (string): Left-leg standing classification from hip-knee-ankle bone-length geometry: standing, non_standing, squat, or uncertain_standing/uncertain_squat.
• stand_right (string): Right-leg equivalent of stand_left.
• percent_stand_left (numeric): Confidence percentage backing the stand_left classification.
• percent_stand_right (numeric): Confidence percentage backing the stand_right classification.
• sit_left (string): Left-leg sitting classification: sitting, non_sitting, or uncertain_sitting.
• sit_right (string): Right-leg equivalent of sit_left.
• percent_sit_left (numeric): Confidence percentage backing the sit_left classification.
• percent_sit_right (numeric): Confidence percentage backing the sit_right classification.
• lie_left (string): Left-side lying classification: lying or non_lying.
• lie_right (string): Right-side equivalent of lie_left.