Pedestrian Intention Forecasting for Autonomous Driving

Pedestrian intention forecasting is a critical task in autonomous driving and intelligent transportation systems, enabling vehicles to anticipate human movement and react accordingly. Understanding pedestrian behavior requires analyzing complex visual and contextual cues to predict future actions. Traditional methods, including rule-based and physics-based models, often struggle to account for the inherent uncertainty and variability in human motion. In contrast, recent advances in deep learning-especially Convolutional LSTMs (Conv-LSTMs) and Vision-Language Models (VLMs)—have shown remarkable performance in capturing both spatial and semantic information from visual data such as images and videos. This thesis focuses on the implementation of two distinct architectures for pedestrian intention forecasting: the first one based on Conv-DNNs and the another one leveraging VLMs for visual feature extraction and context-aware predictions. An open-source dataset, PIE (Pedestrian Intention Estimation), is utilized for training and evaluation, encompassing diverse urban scenarios to enhance model generalization. The two approaches includes data preprocessing, model architecture design, training strategies, and performance evaluation using standard benchmarks. Experimental results demonstrate the performances of each approach separately, highlighting their strengths and limitations. The findings of this thesis seek to contribute to the field of autonomous driving by exploring robust models for pedestrian intention forecasting, with the goal of supporting the development of more responsive and safer autonomous systems.