Multimodal Trajectory Prediction with Intention Recognition and Learnable Path Queries

Trajectory prediction is a critical component of autonomous driving systems. However, it is a highly challenging task due to the complexity of environmental contexts and the multimodal nature of behavioral patterns. To address this issue, we propose an LSTM-based trajectory prediction framework that adopts an encoder-decoder structure. In the encoder, we employ feature-wise and agent-wise attention mechanisms to enhance the LSTM's ability to extract contextual information. Moreover, unlike traditional methods that decode multimodal trajectories using a one-to-many framework based on the features output by the encoder, we propose a novel many-to-many trajectory decoding framework which guides the model in generating diverse trajectories by combining different lane change intentions with learnable path queries. Compared to conventional approaches, our framework provides stronger interpretability, enabling a deeper understanding of the vehicle's complex multimodal behaviors. We conducted experimental validation on the highD dataset, and the results demonstrate that our multimodal trajectory prediction framework outperforms the baseline in terms of prediction accuracy. The open-source code can be found at https://github.com/Yang-JL-bit/trajectory-prediction-with-intention-and-path-queries.git.