Multi-Agent Interaction-Aware Trajectory Prediction based on Behavior Intention for Autonomous Vehicle in Urban Scenarios

• Ano de defesa: 2024
• Autor(a) principal: Gomes, Iago Pacheco
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Autonomous vehicles; Behavior intention prediction; Driving styles; Estilo de direção; Interação; Interaction; Predição de intenção de comportamento; Predição de trajetória; Trajectory prediction; Veículos autônomos
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-02122024-095408/

Autonomous vehicles have the potential to transform urban transport by increasing efficiency, accessibility, and safety while reducing environmental impact. These vehicles use various components to understand the external environment, assess their own state, and interact with other traffic participants. To navigate safely, they employ algorithms for detecting, classifying, and avoiding obstacles. However, merely detecting the position of an obstacle is insufficient for ensuring safety in dynamic urban traffic. Therefore, behavior or intention prediction, and trajectory prediction of these agents are essential. These capabilities enable decision-making and path-planning algorithms to anticipate possible collisions or dangerous situations by considering likely scenarios. The trajectory prediction field encompasses approaches that consider agents motion equations, maneuver intentions, and interactions among traffic participants. Modeling these interactions is particularly challenging due to the complexity of factors influencing each drivers actions, such as psychological factors, driving experience, traffic rules, safety consid- erations, and the actions of surrounding drivers. This dissertation investigates and proposes a novel Multi-Agent Interaction-Aware Trajectory Prediction (MAIATP) framework comprising five components: Road Geometry Modeling, which uses Birds Eye View images to represent global and local features; Driving Style Recognition, that classifies drivers as calm, moder- ate, or aggressive using an Interval Type-2 Fuzzy Inference System with Fuzzy Multi-Experts Decision-Making; Interaction Modeling, that employs a novel graph neural network called the Graph Mixture of Experts Attention Network (GMEAN), which uses prior behavior estimates to weight attention scores; Multi-Agent Interaction-Aware Behavior Intention Prediction, that estimates the lateral and longitudinal behaviors of vehicles and pedestrians, considering their interactions; and, finally, the Trajectory Prediction module, which uses a Conditional Variational Autoencoder (CVAE) to explicitly model the conditional variables inherent to agent behavior in traffic scenarios. The theoretical exploration and experimental validation of these components highlight the importance of interaction in predicting behavior intention and trajectory. The results also demonstrate the advantages of explicitly modeling conditional variables. Finally, this dissertation also addresses the challenges of multimodal prediction, including the imbalanced nature of datasets and the complexity of developing a multi-agent trajectory prediction model for heterogeneous traffic scenarios.