Markerless Human Motion Estimation and Evaluation from Single Video

Malekian, Leila (2024) Markerless Human Motion Estimation and Evaluation from Single Video. Doctoral thesis, University of East Anglia.

Preview

PDF Download (109MB) | Preview

Abstract

Automatic analysis and interpretation of human motion is one of the essential challenges in the field of computer vision and has been the focus of research efforts for many decades. The large quantity of research in this field is motivated by its potential applications in a wide variety of areas and make it interesting across many disciplines and research communities.

One of the main parameters of human motion research projects is the type of input data that is used for processing. Different input modalities include single-view (monocular) video cameras, Kinect depth sensors, RGB multi-camera systems and optical motion capture. The single view video camera is the most commonly used due to it being more affordable and non-specialist and its data being abundant online as compared to the other modalities. However, motions that are occluded from the single camera view, for example, due to self-occlusion, are difficult to recover. In addition to these sources of inaccuracy, the choice of the 3D human model is also important in the captured motion quality. Therefore, we adopted an adaptive shape modeling method called Skinned Multi-Person Linear Model (SMPL) which can make both joint rotations and positions available.

In my PhD research, I propose a machine learning-based method that is used in post-processing to reconstruct the incorrect motions that are caused by self-occlusion. The post-processing network is trained on a data set acquired from different subjects doing 30 different basic exercise motions that include self-occlusion. The collected data comprise single video camera footage and optical motion capture data as the ground truth. To correctly reconstruct the occluded motion, action recognition information is used to select a machine learning model that is trained on the specific motion. The performance of predictive and non-predictive networks are compared to each other and also with the state of the art in human motion estimation. The results show reduction of the overall pose error and the pose error for selected body parts with a large degree of self-occlusion.

Additionally, I have also investigated human motion evaluation which is concerned with how well a specific motion is performed. I have used both classic machine learning with feature extraction and deep learning for this purpose. An improved processing pipeline, feature selection and new machine learning models are used to improve the accuracy of the human motion evaluation compared to the state of the art and baseline methods that are using the same motion evaluation data set.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Chris White
Date Deposited:	07 Jan 2025 09:43
Last Modified:	07 Jan 2025 09:43
URI:	https://ueaeprints.uea.ac.uk/id/eprint/98111
DOI:

Downloads

Actions (login required)

View Item