Human Motion Prediction and Activity Recognition using Pose Estimation
Loading...
Date
item.page.authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Human motion prediction and activity recognition using 3D skeleton-based pose have become
newlinecritical areas of research in computer vision and machine learning, offering significant
newlineadvancements in applications such as surveillance, healthcare, sport and human-computer interaction.
newlineThis study explores state-of-the-art techniques for accurately capturing and interpreting
newlinehuman movement through 3D skeleton data, which provides a robust and detailed
newlinerepresentation of human posture and dynamics. The thesis includes a novel contribution to
newlinehuman motion prediction by developing a Dynamic Multiscale Spatiotemporal Graph Recurrent
newlineNeural Network, which extracts both spatial and temporal features. It also introduces
newlinean Encoder-Decoder Attention Model that incorporates spatial and temporal attention mechanisms
newlineand integrates a Temporal Transformer with a Spatial Graph Convolutional Network
newlinefor improving motion prediction on segmented skeleton body parts. These approaches aim to
newlinepredict human motion in a robust and efficient manner. Several experiments are performed on
newlinethe three datasets Human3.6M, CMU Mocap, and AMASS as discussed in the contributing
newlinechapters. The results demonstrate notable improvements, with the study achieving a significant
newlineperformance boost over the existing baseline on the Human3.6M datasets by 11.95% and 7.74%
newlineof average mean angle errors (avg MAE) for short and long-term motion prediction, respectively.
newlineSimilarly, with CMU Mocap datasets, the DMST-GRNN model predicts future posture
newlinemore accurately than the previous best known approaches by 2.77% and 5.51% of average mean
newlineangle errors (avg MAE) for short and long-term motion prediction, respectively. Additionally,
newlinethe study also predicts human motion more accurately than the previous best methods with
newlinethe AMASS dataset by 2.44%, 1%, 1.25% and 1.3% in terms of Euler angle error, joint angle
newlineerror, positional error, and percentage of correct key point, respectively.
newline