Synthetic Dataset Creation for Fall Detection From Video

This project aimed to create a video dataset which could help train machine learning models to detect instances of people falling in a home setting.

Motion capture was utilized to capture realistic human motion involving different types of falls or non-falls, and real-time rendering technologies were utilized to create videos of digital characters performing these motions in virtual environments. Tools were developed and workflows were established to help produce a large number of rendered videos efficiently, and a video dataset with accompanying metadata was delivered to the partner for their use in training and validating their fall detection models.