Pose-Based Pedestrian Crossing Intention Prediction with Recurrent Graph Convolutions and Explainability-Driven Data Augmentation

This paper presents three contributions to the field of autonomous driving. Firstly, we present a Graph Neural Network designed for the pedestrian crossing intention prediction. The proposed architecture significantly outperforms the state-of-the-art methods in terms of f1-score, precision, and reca...

Descripción completa

Detalles Bibliográficos
Autor: García Romera, Abel
Tipo de recurso: tesis de maestría
Fecha de publicación:2023
País:España
Institución:Universitat Autònoma de Barcelona
Repositorio:Dipòsit Digital de Documents de la UAB
Idioma:inglés
OAI Identifier:oai:ddd.uab.cat:293067
Acceso en línea:https://ddd.uab.cat/record/293067
Access Level:acceso abierto
Palabra clave:Advanced Driver Assistance Systems
Data augmentation
Explainability
Graph Neural Networks
Pedestrian Crossing Intention
Descripción
Sumario:This paper presents three contributions to the field of autonomous driving. Firstly, we present a Graph Neural Network designed for the pedestrian crossing intention prediction. The proposed architecture significantly outperforms the state-of-the-art methods in terms of f1-score, precision, and recall. The input consists of the skeletons of the pedestrians. It extracts features from the skeletons using a Recurrent Graph Convolutional Layer that captures both the spatial and temporal dependencies. Secondly, an explainability analysis has been performed using our implementation of Grad-Cam. Finally, based on the results of the explainability, we have proposed a data augmentation method that generates new skeletons.