Query-enhanced motion transformer with dilated static query and bridged dynamic query

Motion forecasting is integral to autonomous driving, as it aims to anticipate future trajectory of traffic agents within the complicated environment by comprehending their various motion intentions. Recent effective query-based methods have focused on designing static queries and learning dynamic queries through cascaded Transformer decoders to represent motion multimodality. However, static queries are usually generated by clustering the goal of motion samples, while they suffer from out-of-distribution limitation and thus deviate from the boundary samples, leading to a suboptimal prediction. Additionally, dynamic queries, affected by the sequential connections between decoder layers, lack communication between distant decoders, leading to inconsistent prediction. In this paper, we propose a simple yet effective Query-Enhanced Motion TRansformer (QE-MTR) for motion forecasting. This approach includes both Dilated Static Query (DSQ) and Bridged Dynamic Query (BDQ) to enhance the representation of motion multimodality. Specifically, the DSQ expands the valid range of motion samples from their future sequence to complete sequence, providing more diverse motion patterns for query generation and effectively dilating the coverage of static queries. The BDQ introduces cross connection across decoders instead of direct connections, promoting harmonious dynamic optimization and improving the effectiveness of final dynamic queries. QE-MTR achieves state-of-the-art performance on the Waymo Open Motion Dataset, earning 2nd place finish in the 2023 Waymo Motion Prediction Challenge.