Pose Estimation and Analysis for Powerlifting

Notice

********************************************************************************
*    Pose Estimation and Analysis for Powerlifting                             *
*    Final year dissertation project                                           *
*    Kyle Moir <katm2000@hw.ac.uk>                                             *
*    2023-2024                                                                 *
*                                                                              *
*    NEITHER THIS APPLICATION NOR ITS DEVELOPERS ARE LIABLE FOR ANY INJURY     *
*    CAUSED AS A RESULT OF THIS PROGRAM.                                       *
*                                                                              *
*    Always take care when practising powerlifting.                            *
********************************************************************************

This project uses Mediapipe BlazePose human pose estimation to analyse the barbell squat and deadlift powerlifts. Bench press has been excluded due to the practical challenges faced when using a 2D camera to capture footage (obstructed joints).

Usage

Dependencies

• Ubuntu 20.04 or later is recommended.
• Python 3 with PIP.

Install required packages

The required packages can be installed via pip3 install -r requirements.txt.

Running

The program runs from the command line python3 main.py.

Program Arguments

- `--help`
Print (this) help message

- `--credits`
Print software credits

- `--input | -i <filepath>`
Video file input

- `--movement | -m <squat/deadlift>`
Override movement detection

- `--verbose | -v`
Print more verbose info

- `--dry`
Dry run. Estimate pose and don't do any evaluation.

Guide

The program should be provided with an mp4 video file containing footage of either a barbell squat or deadlift. Ideally the video has the subject fully in frame against a contrasting background. The program is tailored to evaluating a single rep, though some of the on-screen metrics will continue to be reported after a single rep is finished.

Example input data is provided in the input data folder.

The exercise will be automatically identified by the program. If you have trouble with this, try the --movement <squat/deadlift> argument to override this automatic identification.

Analysing a video typically processes at half real time, ideally input videos are around 10 seconds or less.

Metrics

Squat

1. Squat depth The average distance between each ankle to between the hips. Higher is better.

2. Knee balance The average of the horizontal distance between the knee and shoulder on each side of the body. This metric gives us an indication of the symmetry of the lift. Lower is better.

3. Time The duration of the squat. Lower is better.

4. Hips below knees True/false value determining if the hips of the lifter were below the knees. True is ideal.

Deadlift

1. Complete lift True/false value determining if the bar pass the shins of the lifter and down again. True is ideal.

2. Downward movement The number of seconds where the bar is lower than the previous second, before the apex of the lift. note: this metric is not reliable due to the 'noisy' nature of pose estimation. Lower is better.

3. Time The duration of the deadlift. Lower is better.

Output

When finished, the program will print to the console a list of metrics gathered, and a file name_output.mp4 (from the input name.mp4).

• This video will have the landmarks of the subject rendered as red nodes, with green edges connecting them.
• The top left shows some of the metrics as they are being gathered.
• At the left wrist of the lifter is the 'drop count'. This metric determines if the bar has dropped before the apex of the lift.
• When analysing a barbell squat, two angles of the lat and knee will be rendered. The greener the text, the acuter the angle.