Behavior-Pred: A semantic-enhanced trajectory pre-training framework for motion forecasting

Abstract

Predicting the future movements of dynamic traffic agents is crucial for autonomous systems. Effectively understanding the behavioral patterns of traffic agents is key to accurately predicting their future movements.

Inspired by the success of the pre-training and fine-tuning paradigm in artificial intelligence, we develop a semantic-enhanced trajectory pre-training framework for motion forecasting in the autonomous driving domain, named Behavior-Pred. In detail, we design two kinds of tasks during the pre-training phase: fine-grained reconstruction and coarse-grained contrastive tasks, to learn a better representation of both historical and future behaviors, as well as their pattern consistency. In fine-grained reconstruction learning, we utilize a time-dimensional masking strategy based on the timestep level, which reserves historical and future patterns compared to agent-based masking. In coarse-grained contrastive learning, we design a similarity-based loss function to grasp the relationship/consistency between history patterns and the future. Overall, Behavior-Pred learns more comprehensive behavioral semantics via multi-granularity pre-training tasks. Experimental results demonstrate that our framework outperforms various baselines.