Studying and Improving Motion Indeterminacy Diffusion for Stochastic Trajectory Prediction

Human trajectory prediction is a fast growing subject in the Computer Vision field; the study of this problem is acquiring more and more importance due to its practical implications, among which autonomous driving and crowd surveillance are the most cited. Many different points of view are being analyzed in order to improve the precision and the correctness of the models; for instance, different data can be exploited to better understand how humans behave, and different mathematical frameworks can be used to improve the predictions. One big acknowledgment in the field is that this is a challenging task due to the inherent indeterminacy human motion brings: free will and unpredictability are, in fact, intrinsic to human behavior. The authors of the paper on which this work is based try to solve this very issue: their idea, based on the concept of diffusion models, is to simulate this indeterminacy by applying random noise and subsequently try to learn how to remove it from the available data in order to obtain the most probable trajectories. Their work, however, is based on an outdated version of diffusion models; many advancements have been made in the meantime, thus the idea for this work: we try to modify the model by applying the improvements, and thus try to improve its performance. In addition to this, we tried to expand the model by implementing a goal module, already developed and currently in improvement by the VIMP group at University of Padua, in order to include additional information and, therefore, to enhance the model capabilities.