Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Skeleton detection from images and videos, oriented for ski technique analysis

Notifications You must be signed in to change notification settings

TuronLab/PoseDetection

Repository files navigation

Summary

This is a quick attempt to calculate the pose of a single person in a video or image. All the notebooks in this repository has been focused on technique analysis of single cross country skiers. So, in this repository you will find 4 notebooks for different tasks and strategies to study the pose of the skier. So, in this repository you will find a notebook focused in calculate the pose of a single skier in a picture (ImagePose notebook). There is another notebook that perform the same strategy in the previous notebook, but it will calculate it for every frame of a video, and will record it (VideoPose notebook). However, this strategy tends to fail when skier is small in the frame (skier is far), so, with the help of a Yolov3, you can find every person in frames, and cut the image to zoom in skier (example of yolo detection in VideoPersonDetection notebook). In this way, you can find a solution in VideoOpenPose&PersonDetection notebook, that calculates the pose of skier doing zoom in skier when he is far, and without zoom when he is near. If Yolo decets more than one person, we will focus on the biggest one, because video it's supposed to be focused on the skier that is being studied.

Because of the quality of single frames, just for calculate the pose, you can choose an option to blurr vertical lines between frames with a gaussian filter (5,5) in the video notebooks. This blurring won't be in final result, and it uses to improve the result.

Here, we have an example of how it works for a video of a skier going uphill.

alt text

Models for Pose Detection

For this approach, I have used the pre-trained models weights files using the scripts provided here. This models, based in the architecture described in Zhe Cao, et al. (2016), has been trained for two different datasets: COCO, and MPI. The model trained for COCO, produces 18 points, and MPII, produces 15. For this application, I have decided to use MPII model, because it has cross country skiers in the dataset.

Model for Person Detection

For the recognition of people in images, I have used the pre-trained model YOLOv3-spp that you can download here.

About

Skeleton detection from images and videos, oriented for ski technique analysis

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published