EECS 442 - Introduction to Computer Vision Final Project

University of Michigan - Shruti Ambekar, Jonas Hirshland {ashruti, jhirshey}@umich.edu

Estimating Figure Trajectory using OpenPose

Report

Setup

Carnegie Mellon OpenPose

We built our project using processed frames from CMU's OpenPose.

The instruction steps for all operating systems should be in there. You can clone this repository in the following path in OpenPose after you have configured using CMake. OpenPose_directory/build/examples/tutorial_api_python/. The only file that needs to be moved from this repository to that repository is the multiple_frames.py file. It will not run correctly if it remains in this project repository.

Our Files

• parse_video.py takes a command line arguments --filename, --frame_rate, and --output which takes in a .MOV file and outputs JPGs after skipping as many frames as put in the frame rate argument.

• multiple_frames.py is a slightly tweaked 01_keypoints_from_images.py from OpenPose that correctly stores the keypoints in .npz files so we can reuse them instead of having to rerun OpenPose.

• estimate[x].py are the different estimations we used, using linear, and other methods.

• figure[x].py are some pyplot figures created to help illustrate our methodology in the report.

• raw_videos/ are the videos we tested our estimations on.

• processed_videos/ are the processed videos after running parse_video and multiple_frames scripts on our input data.