MCIP: Multi-Stream Network for Pedestrian Crossing Intention Prediction

Je-Seok Ham, Kangmin Bae, Jinyoung Moon

Abstract: Predicting the crossing intention of pedestrian is an essential task for autonomous driving systems. Whether or not a pedestrian will cross a crosswalk is a significantly inevitable skills for safety driving. Although many datasets and models are proposed to precisely predict the intention of pedestrian, they lack the ability to integrate different types of information. Therefore, we propose a Multi-Stream Network for Pedestrian Crossing Intention Prediction (MCIP) based on our novel optimal merging method. The proposed method consists of integration modules that takes two visual and three non-visual elements as an input. We achieved state-of-the-art performance on accuracy of pedestrian crossing intention, F1-score, and AUC with both public standard pedestrian datasets, PIE and JAAD. Furthermore, we compared the performance of our MCIP with other networks quantitatively by visualizing the intention of the pedestrian. Lastly, we performed ablation studies to observe the effectiveness of our multi-stream methods.