DEVELOPMENT OF GRAPH NEURAL NETWORK APPROACH FOR PEDESTRIAN TRAJECTORY PREDICTION

Abstract

Pedestrian trajectory prediction is a fundamental problem in autonomous systems, robotics, and self-driving vehicles, where accurately forecasting future human motion is critical for safe navigation and interaction. The task is particularly challenging due to the highly dynamic, socially interactive, and inherently multimodal nature of pedes trian behavior. This thesis first presents a comprehensive study of the pedestrian trajectory pre diction domain, ranging from physics-based models and classical machine learning techniques to modern deep learning and multimodal generative frameworks. Existing methods are analyzed using commonly adopted evaluation metrics such as Average Displacement Error (ADE) and Final Displacement Error (FDE). The review high lights several persistent limitations in current systems, including poor generalization across scenes, limited robustness to irregular crowd behavior, computational complex ity, and challenges in modeling diverse future intentions. Motivated by these observations, this thesis proposes two novel deep learning ar chitectures for multimodal pedestrian trajectory forecasting. The first proposed model, SWIFT-Net (Slot-guided Wavelet Interaction Fusion for Trajectories), introduces a hi erarchical framework for capturing both local and global interaction patterns. Sep arate spatial and temporal interaction graphs are constructed and decomposed using wavelet-based multi-scale feature extraction, enabling the model to learn short-range and long-range dependencies simultaneously. An adaptive fusion mechanism together with slot attention disentangles distinct behavioral patterns within crowds, allowing the model to generate multiple socially plausible future trajectories. The second pro posed model, TRACE-Net (Trajectory Reasoning via Adaptive Competitive Encod ings), focuses on adaptively modeling irregular crowd dynamics through deformable spatio-temporal reasoning. The architecture employs asymmetric and deformable 3D convolutions to dynamically adjust receptive fields based on interaction patterns. In addition, a Competitive Attention mechanism is introduced to separate latent multi modal intentions into distinct behavioral encodings. These representations are further processed using an Attention-Gated Temporal Decoder to generate diverse and socially consistent future trajectories. Extensive experiments conducted on the ETH and UCY benchmark datasetsdemon strate the effectiveness of the proposed approaches in improving trajectory prediction accuracy while maintaining diversity and social plausibility. Overall, this thesis con tributes both a detailed understanding of the current state of pedestrian trajectory pre diction research and novel architectures designed to address key limitations in existing methods.