Learning Intention-Behavior Consistency for Pedestrian Crossing Prediction

Abstract

Pedestrians are the vulnerable group in road traffic, so understanding the behavior of pedestrians and accurately predicting their crossing intentions is of great significance for traffic safety. Most existing methods treat the prediction task as a binary classification problem. However, such traditional methods usually can only establish a deterministic mapping between the environment and the pedestrian's state, and cannot fully model the multimodality of the pedestrian's intentions in highly dynamic scenarios. In recent years, researchers have introduced conditional generative models to capture these unobserved uncertainties by introducing latent variables, thereby enabling the modeling of complex intention distributions. Nevertheless, in highly dynamic traffic scenarios, traditional multimodal conditional generation modules still have significant limitations. To address these problems, we propose a dynamic consistency constraint module, namely MAA-IAC, which aligns the action and intention modalities and improves the model performance gradually. To evaluate the proposed framework, we conducted extensive comparative experiments studies on the JAAD benchmark dataset. The results show that introducing dynamic intention-behavior consistency constraints can significantly improve the prediction performance of the model. Specifically, our proposed method boosts the baseline Accuracy from 0.7945 to 0.8754, while also achieving notable improvements in other key metrics such as F1-score and AUC.