Hi-SCL: Fighting long-tailed challenges in trajectory prediction with hierarchical wave-semantic contrastive learning

Predicting the future trajectories of traffic agents is a pivotal aspect in achieving collision-free driving for autonomous vehicles. Although the overall accuracy of existing prediction methods appears promising, most of them overlook the long-tailed challenge in trajectory prediction. They tend to excuse or overlook the disastrous performance in rare yet safety-critical tail events. This paper puts forward a novel framework called hierarchical wave-semantic contrastive learning (Hi-SCL), which attempts to fight the long-tailed challenge in the trajectory prediction task. Our approach innovatively represents each traffic scene as “waves”, and implicitly models traffic multi-stream interactions through wave superposition at both local and global levels. This pioneering incorporation of the wave concept enhances the in-depth comprehension of the traffic scene. On this basis, we introduce the feature hierarchical reshaping method, empowering our network to cope with formidable infrequent cases effectively. This module maintains a collection of feature-enhanced hierarchical prototypes, dynamically steering trajectory samples closer or pushing them farther away in an unsupervised learning setup. Extensive experiments on real-world datasets validate Hi-SCL’s robust overall prediction performance and its effectiveness in addressing long-tailed challenges. Compared to several baseline models, Hi-SCL demonstrates remarkable improvements in general predictive accuracy, with long-term prediction error reductions ranging from 14% to 54% for minADE and 27% to 79% for minFDE. The outcomes of long-tailed experiments further underscore the capacity of Hi-SCL, offering accuracy gains ranging from 2% to 17% in tailed samples. The thorough empirical analyses confirm Hi-SCL’s exceptional capability of wave-semantic representation learning and its effectiveness in reshaping the feature space via hierarchical contrastive learning mechanisms. The proposed new paradigm paves the way for substantial advancements in trajectory prediction, especially in overcoming long-tailed issues, bringing us closer to realizing safer autonomous driving systems.