Predicting Pedestrian Intent: Evaluating Deep Learning with LSTM, CNN, and Transformers

Pedestrian Intent Prediction (PIP) plays a key role in the safety of autonomous driving systems by predicting whether a pedestrian intends to cross a street. This paper presents a comprehensive analysis of multiple deep learning architectures for PIP, including Long Short-Term Memory (LSTM), convolutional models, and Transformer-based networks. The study evaluates model performances using the PSI 2.0 dataset, which provides annotated videos of pedestrian behaviours. We focus on enhancing prediction accuracy while tackling the challenges of class imbalance. Our experiments show a significant improvement in performance when switching to Transformer models and incorporating techniques such as class weighting and focal loss.