Multi-Person Pose Tracking Using Dynamically Gated Similarities

Human pose tracking is a crucial task for many applications, such as human-robot interaction or video surveillance. Most current methods use a tracking-by-detection approach, which has shown to be very successful. These methods either use the location and movement of the detected humans or their visual appearance to track them. In recent years, some models started using heuristic approaches to combine both the pose and the appearance information. Even these newer models are still using fixed weights for these combinations. Therefore, the weights are dataset-dependent and need to be tuned for each dataset. In this thesis, we propose a new model that uses dynamically learned weights to combine the pose, appearance, and or other information. The so-called Dynamically-Gated Similarities (DGS) model is highly modular and customizable. Weights can be used and trained mostly dataset independent. Additionally, the weights are computed individually for every detection, allowing the model to adapt more freely to the current situation. The DGS model is evaluated on the PoseTrack21 and the DanceTrack datasets. The proposed algorithm does not outperform current state-of-the-art models due to the simplicity of the used parts, but it is shown that dynamically learned weights can increase the tracking performance of the basic modules slightly. The implementation and results will be made publicly available at: https://github.com/bmmtstb/dynamically-gated-similarities. The modules are structured in a way that they can be easily replaced by more complex models, which should increase the overall performance of the dynamically gated similarities model. Thus, this thesis only provides a proof of concept and baseline models for the dynamically gated similarities algorithm.