Multi-Keypoint Tracking on Synthetic Crowd Data

Bachelor Thesis, Oliver Becker, 2020

Due to the great advances in machine learning, many complex problems now become tractable. One such problem is pose estimation, which has attracted a lot of attention recently. Image-based pose estimation aims at determining the pose of humans in an input image. In order to obtain more information than a single pose per time step, pose tracking is necessary. Pose tracking uses the information about the individual poses from the pose estimation in order to follow the poses’ and hence the human’s motion. This work first gives an overview of pose tracking, its applications in the recognition of human behaviours, and its challenges. An application we focus on specifically is crowd tracking. Due to its connection to pose tracking, we also address the pose estimator we employ. The main focus of this work is the investigation of a pose tracking algorithm based on a Bayesian filter, more specifically a particle filter. For that purpose, we go into detail on the theory on particle filters and its application in pose tracking algorithms. For the prediction step of the particle filter, we implement a probabilistic learning algorithm for pose prediction based on pose sequences. We finally examine the performance of our algorithm on synthetic datasets.