Transformer based stochastic approach to pedestrian trajectory forecasting

Pedestrian trajectory forecasting is an active field of research in which the goal is to predict the path of an agent over a period of time, given some previously observed past motion history. This predicted period of time can be long or short, leading to long-term forecasting or short-term forecasting. In this thesis, we deal with short-term trajectory forecasting which consists in a forecasting window of about five seconds in the future. Having a model capable of these kinds of predictions is useful in many applications such as for service robots, autonomous driving or anomaly detection, where predicting a possible outcome is necessary to plan future moves. A great challenge in modelling human motion behaviour is the fact that it is not deterministic by nature. An agent movement and trajectory is determined by many forces but also by the own will of the agent, which is not predictable with absolute certainty. In fact, multiple future trajectories may be plausible in certain situations such as in front of a fork in the road, or to avoid a collision, and even if the intent and destination of the agent are known, there are possibly infinite trajectories which can be possible. This stochastic and multimodal behaviour is the main focus of this thesis. With the advent of the Transformer model for Natural Language Processing, more and more Transformer based models have been proposed for sequence modelling problems. Motivated by their good performance with respect to Recurrent Neural Networks also for this task, the experiments of this thesis have been conducted using Transformer based architectures. In the first part of the experiments, the used model from literature, Transformer TF, is introduced and different preliminary experiments are carried out on the model to improve its performances, such as lowering model size, changing the inputs and perform a data augmentation technique which is proven to be beneficial. Then, starting from the previous deterministic model, a proposal is made to take into account the problem stochasticity. The proposed model, Transformer GAN, is a generative model capable of multiple predictions thanks to a sampled random latent vector. To efficiently generate multiple and multimodal predictions, a sampling refinement technique is then introduced which drastically improves the performances of the stochastic approach used. The results of the Transformer GAN after the sampling refinement method reach state-of-the-art performances with respect to other deep learning models for trajectory forecasting that do not use any social nor map information.