 |
|
| Pose estimation example using OpenPose. | Example bicycle crash |
Traditional methods to analyse the motion of bodies, such as humans or vehicles, rely on Motion Capture systems or plenty of sensors. These methods offer great precision and reliability, however, they are only available in specialised laboratories or instrumented vehicles, which limits the analysis of 'in-the-wild' data. For this reason, during the last years, and supported by the constant performance increase of computer vision techniques, the scientific community have been developing tools to extract motion data from normal videos. This method is called 'Pose estimation' and has received great attention in the last ten years, mainly focused on human motion. However, there is no publicly available implementationof this methodology for estimating the motion of a bicycle. Therefore, we propose a project to implement similar methodologies into bicycle motion analysis, which will allow researchers to analyse bicycles in-the-wild and enhance their understanding.
Objective: Extract bicycle's kinematic data from monocular videos and compare the performance of different training approaches.
Suggested approach
To carry this project, the suggested approach consists in a three-stage pipeline, based on KinePose:
- Use computer vision algorithms to detect keypoints on the bicycle.
- Convert 2D keypoints into 3D point cloud with different methodologies.
- Compare the performance of the different methodologies.
Known methodologies
From MotionBert, it is known that three different methodologies to go from 2D keypoints to 3D cloud of points are:
- Heatmap from depth estimation.
- Train the algorithm with motion capture data.
- Train the algorithm with synthetic data.
PDF version: Proposal
Don't hesitate to reach out at b.gonzaleztoledo@tudelft.nl if you have more questions.
How To Apply
If you would like to apply for this project, please send an approximately half page letter explaining your motivations and interest in the lab and project, CV or resume, a list of courses you've taken, the name of your MSc track, and any other relevant information to j.k.moore@tudelft.nl.