Visual–Motion–Interaction-Guided Pedestrian Intention Prediction Framework

The capability to comprehend the intention of pedestrians on the road is one of the most crucial skills that the current autonomous vehicles (AVs) are striving for, to become fully autonomous. In recent years, multimodal methods have gained traction employing trajectory, appearance, and context for predicting pedestrian crossing intention. However, most existing research works still lag rich feature representational ability in a multimodal scenario, restricting their performance. Moreover, less emphasis is put on pedestrian interactions with the surroundings for predicting short-term pedestrian intention in a challenging ego-centric vision. To address these challenges, an efficient visual–motion–interaction-guided (VMI) intention prediction framework has been proposed. This framework comprises visual encoder (VE), motion encoder (ME), and interaction encoder (IE) to capture rich multimodal features of the pedestrian and its interactions with the surroundings, followed by temporal attention and adaptive fusion (AF) module (AFM) to integrate these multimodal features efficiently. The proposed framework outperforms several SOTA on benchmark datasets: Pedestrian Intention Estimation (PIE)/Joint Attention in Autonomous Driving (JAAD) with accuracy, AUC, ${F}1$ -score, precision, and recall as 0.92/0.89, 0.91/0.90, 0.87/0.81, 0.86/0.79, and 0.88/0.83, respectively. Furthermore, extensive experiments are carried out to investigate different fusion architectures and design parameters of all encoders. The proposed VMI framework predicts pedestrian crossing intention 2.5 s ahead of the crossing event. Code is available at: https://github.com/neha013/VMI.git .