Human Pose Data Capture System for Training Fitness Apps

December 22, 2021

The raising interest about human behaviour recognition on fields like robotics, fitness and security led to a growing attention of the academic community to computer vision problems and image analysis. In the past few years deep learning techniques, like neural networks, have been used extensively to improve the accuracy detection of the human keypoints, which are essential to describe human behaviour, poses and gestures on images. These methods need datasets and their accuracy is deeply tied with the quantity and quality of the datasets used for training. A dataset consists of not only the images but also the labels with the ground-truth localization of each keypoint of the human body, thus their creation is a very time consuming task and prone to errors. For the specific case of fitness images they tend to be more challenging since they can be fast paced, have occlusions and unusual body positions. In this work we propose a semi-supervised calibration method, with a motion capture system and smartphone cameras, to create datasets with high accuracy and low effort for training, validating or evaluating human pose estimation algorithms.

Authors in the community:

• 

  B

  Bruno Miguel Ribeiro Marta

  ist156673

Supervisors of this institution:

• 

  A

  Alexandre José Malheiro Bernardino

  ist13761

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico