Pedestrian Motion Prediction Using Deep Learning

Pedro Miguel Gustavo Bilro

Key information

Authors:

Pedro Miguel Gustavo Bilro (Pedro Miguel Gustavo Bilro)

Supervisors:

Jorge Dos Santos Salvador Marques (Jorge Dos Santos Salvador Marques); Ana Catarina Fidalgo Barata (Ana Catarina Fidalgo Barata)

Published in

11/22/2021

Abstract

Pedestrian motion prediction is a task that is relevant for many kinds of intelligent systems, such as autonomous vehicles, social robots, and automated surveillance systems. It can be a very challenging task, due to the fact that humans can be influenced by a plethora of factors. In recent years, two types of cues have been getting more relevance: the presence of physical obstacles, and the existence of social interactions between pedestrians. While there have been methods that incorporate both types of cues, most require extra information such as video images, which may not be readily available. This thesis proposes a novel method, based on Long Short Term Memory, that takes into account obstacles and neighbouring humans to robustly predict pedestrian trajectories. The presence of obstacles is incorporated by integrating the model with Sparse Motion Fields, that learn regions absent of pedestrian motion in an unsupervised way. The existence of interactions is incorporated with a pooling layer that simulates a field of view for each pedestrian. The proposed method is able to outperform several state-of-the-art models on popular pedestrian datasets. To compare the models, there was the use of standard geometric errors, as well as metrics specifically related to static obstacles and dynamic interactions. The experiments also included visualization of method predictions, to provide a different perspective on the performance of trajectory forecasting models. The model implementation and experiments are publicly available and can be accessed at https://github.com/pedro-mgb/pedestrian-arc-lstm-smf.