MF-BERT: A Siamese Pre-training Framework for Motion Forecasting

Accurately predicting the future motions of traffic agents is essential for autonomous systems. Despite the significant success of existing motion forecasting methods based on supervised learning, they still exhibit two main limitations. First, when annotated data for a scene is limited, these methods often fail to achieve the expected accuracy. Second, they typically rely on complex architectures and extensive prior knowledge to improve performance. To overcome these challenges, we propose MF-BERT, a novel framework that adapts the concept of BERT to motion forecasting, inspired by advancements in the self-supervised pre-training paradigm. During pre-training, we design a siamese sequence modeling task with an asymmetric mask strategy to capture complex behavior patterns of agents. During fine-tuning, the pre-trained representation module initializes the feature encoder of the motion forecasting model, and a multimodal trajectory decoder generates all possible predictions. Experimental results demonstrate the superiority of MF-BERT over state-of-the-art methods.