charlesyao2005/cprcoachrate

CPR-Coach compression count and rate

Chest-compression count and rate for every camera view in CPR-Coach, estimated from its frames: one row per view, 5,664 rows (1,416 videos × 4 cameras). The frames themselves aren't included; they're in CPR-Coach.

These are automatic estimates from pose tracking, not human annotations.

from datasets import load_dataset

ds = load_dataset("charlesyao2005/cprcoachrate", split="train")

Why one row per view

A video's four views don't always show the same moment. In some recording sessions one camera's clips are shifted in time against the other three, as the clock burned into each frame shows:

• CPR_Double_Dataset_S0, ch2, DC15–DC31: about 12 s late on clips of 7–14 s (checked against the clocks for DC29 and DC31).
• CPR_Double_Dataset_S0, ch1, most of DC00–DC14: undercounts the same way (not checked against the clocks).
• CPR_Dataset_S1, ch3, C00–C04: every view counts 23–25 instead of 30, consistent with a shift of a few seconds (not checked against the clocks).

Such a view shows only part of the set, part of the next repetition, or an empty room. So each row counts what its own clip shows; the views aren't combined.

Columns

column	type
path	string	view, e.g. CPR_Dataset_S0/C13_S0/r00/ch0, as in CPR-Coach's ann/*.txt
sample	string	video: path without /chN
subset, class_dir, rep, view	string	the parts of path
frames	int	frames in the clip (25 fps)
duration_s	float	frames / 25
compression_count	int	compressions in the clip (null if the rescuer wasn't tracked)
compression_rate_cpm	float	compressions per minute while compressing (pauses left out)
pose_found	float	fraction of frames in which a person was tracked
unreliable	bool	no count, or fewer than 6 compressions
note	string	why, if unreliable

path is the first field of CPR-Coach's ann/*.txt lines, so the action labels join on it: the line CPR_Dataset_S0/C00_S0/r00/ch0 286 0 describes the row with that path.

Method

1. Pose. RTMPose (rtmlib's "balanced" models: YOLOX-m person detector, RTMPose-m, COCO-17 keypoints) on every frame at half resolution. The rescuer is the person whose shoulders are highest when first found (they kneel over the manikin), then whoever is nearest their last position.
2. Signal. The median vertical position of the rescuer's nose, eyes, ears, shoulders and elbows, each about its own resting height, band-passed to 0.5–4 Hz (30–240 compressions/min).
3. Count. Peaks of downward motion at most 200/min apart, each moving the upper body by at least 0.025 torso lengths and at least 40% of the clip's median peak. A view needs the rescuer in at least half its frames to get a count.
4. Rate. 60 / the mean interval between consecutive compressions, leaving out intervals longer than 1.5× the median, i.e. pauses.

The counting follows the keypoint approach of Zhang et al., "Quality Evaluation Algorithm for Chest Compressions Based on OpenPose Model" (Appl. Sci. 2022, 12(10), 4847), with RTMPose in place of OpenPose.

How reliable

The protocol is 15 or 30 compressions per clip. Of the 5,631 views not marked unreliable, 3,933 count 30, 849 count 15 and 849 count something else. The median rate is 125.4/min (5th–95th percentile 48.9–162.3).

A count other than 15 or 30 isn't necessarily wrong. Known reasons:

• The volunteer did more. Clips where all four views agree on 31 or 32 with evenly spaced compressions; checked frame by frame for CPR_Dataset_S0/C13_S0/r01, which has 31.
• The view is shifted in time (above), so it shows only part of the set.
• The clip ends before the set does. The clips are trimmed tightly around the compressions, and some are shorter than the set at the rescuer's pace.
• Counting errors, usually ±1: a compression in progress at the clip's first or last frame isn't counted, and pauses or irregular movement can hide one.

The 33 unreliable views mostly have the rescuer out of shot (21 are CPR_Double_Dataset_S0 ch2). With nobody else in the frame the tracker can follow the manikin instead, so pose_found stays high even then.

Source

CPR-Coach: Wang et al., arXiv:2309.11718.