Moving Pose

2 minute read

Project’s GitHub Repo

Our Presentation (December 2020)

Members:

  • Andrew Darling
  • Eric Hayes
  • Mehmet Yilmaz

About:

This project is our implementation of the Moving Pose algorithm which was proposed by Mihai Zanfir, Marius Leordeanu, & Cristian Sminchisescu in there paper. The algorithm is used to recognize and understand human actions quickly and accurately using “skeleton” data from a depth sensor.

Notes:

  • Given a skeleton based dataset collected from a depth sensor, the goal is to classify certain human actions using the Moving Pose algorithm as well as provide a simple UI.
  • To achieve this goal, we implemented the Moving Pose algorithm from the paper stated below and the database stated below.
  • This is our Fall 2020 CSCI470 (Introduction to Machine Learning) final Project. CSCI470 is an undergraduate class provided at the Colorado School of Mines. Our team name was: Nestlé.
  • Please view /movingpose/gui/README.md to learn more about the GUI and the hardware(s) used.

Paper Implemented:

  • Title: The Moving Pose: An Efficient 3D Kinematics Descriptor for Low-Latency Action Recognition and Detection
  • Authors: Mihai Zanfir, Marius Leordeanu, & Cristian Sminchisescu.
  • Paper: Zanfir_The_Moving_Pose_2013_ICCV_paper.pdf

Dataset Used:

  • We used the MSR DailyActivity 3D Dataset dataset: Dataset_Source
  • Multiview Action 3D Dataset Action IDs: 3

UI Preview: