MoveNet Multipose Tracking on OpenVINO

Running Google MoveNet Multipose models on OpenVINO.

A convolutional neural network model that runs on RGB images and predicts human joint
locations of several persons (6 max).

WIP: currently only working on CPU (not on GPU nor MYRIAD)

Demo

Full video demo here.

For MoveNet Single Pose, please visit : openvino_movenet

Install

You need OpenVINO (tested on 2021.4) and OpenCV installed on your computer and to clone/download this repository.

Run

Usage:

> python3 MovenetMPOpenvino.py -h
usage: MovenetMPOpenvino.py [-h] [-i INPUT] [--xml XML]
                            [-r {192x192,192x256,256x256,256x320,320x320,480x640,736x1280}]
                            [-t {iou,oks}] [-s SCORE_THRESHOLD] [-o OUTPUT]

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT, --input INPUT
                        Path to video or image file to use as input
                        (default=0)
  --xml XML             Path to an .xml file for model
  -r {192x192,192x256,256x256,256x320,320x320,480x640,736x1280}, --res {192x192,192x256,256x256,256x320,320x320,480x640,736x1280}
  -t {iou,oks}, --tracking {iou,oks}
                        Enable tracking and specify method
  -s SCORE_THRESHOLD, --score_threshold SCORE_THRESHOLD
                        Confidence score (default=0.200000)
  -o OUTPUT, --output OUTPUT
                        Path to output video file

Examples :

To use default webcam camera as input :

python3 MovenetMPOpenvino.py
To specify the model input resolution :

python3 MovenetMPOpenvino.py -r 256x320
To enable tracking, based on Object Keypoint Similarity :

python3 MovenetMPOpenvino.py -t keypoint
To use a file (video or image) as input :

python3 MovenetMPOpenvino.py -i filename

Keypress	Function
Esc	Exit
space	Pause
b	Show/hide bounding boxes
f	Show/hide FPS

Input resolution

The model input resolution (set with the '-r' or '--res' option) has an impact on the inference speed (the higher the resolution, the slower the inference) and on the size of the people that can be detected (the higher the resoltion, the smaller the size).
The test below has been run on a CPU i7700k.

Resolution	FPS	Result
192x256	58.0
256x320	44.1
480x640	14.8
736x1280	4.5

Tracking

The Javascript MoveNet demo code from Google proposes as an option two methods of tracking. For this repository, I have adapted this tracking code in python. You can enable the tracking with the --tracking (or -t) argument of the demo followed by iou or oks which specifies how to calculate the similarity between detections from consecutive frames :

IoU (Intersection over Union) of pose bounding boxes (option iou);
Object Keypoint Similarity (option oks).

Tracking	Result
IoU Tracking
OKS Tracking

In the example above, we can notice several track switching in the IoU output and a track replacement (2 by 6). OKS method is doing a better job, yet it is not perfect: there is a track switching when body 3 is passing in front of body 1.

The models

The MoveNet Multipose v1 source model comes from the Tensorfow Hub: https://tfhub.dev/google/movenet/multipose/lightning/1

The model was converted by PINTO in OpenVINO IR format. Unfortunately, the OpenVINO IR MoveNet model input resolution cannot be changed dynamically, so an arbitrary list of models have been generated, each one with its dedicated input resolution. These models and others (other resolutions or precisions) are also available there: https://github.com/PINTO0309/PINTO_model_zoo/tree/main/137_MoveNet_MultiPose