DMPT: An Efficient Dual-Mamba Framework for Pedestrian Trajectory Prediction

Accurate and efficient pedestrian trajectory prediction is crucial for autonomous driving system, enabling safe and timely decision-making in complex urban environments. Despite much progress made in deep learning models, existing methods still struggle to simultaneously model intricate spatiotemporal interactions and inherent uncertainty in future trajectories, while maintaining computational efficiency. To address these limitations, we propose an efficient Dual-Mamba framework (DMPT), integrating intrinsic motion modeling and extrinsic social interactions into a unified encoder-decoder Mamba-based architecture. Specifically, we construct motion hypotheses for representative motion patterns from observed trajectories and employ a bidirectional Motion-Mamba encoder (BiMME) with learnable positional embeddings to thoroughly capture inter-pattern dependencies in both forward and backward directions. A classification head is then used to produce probabilities over motion patterns for selection and multimodal prediction. Conditioned on the selected motion hypotheses, a social Cross-Mamba decoder (SCroMD) incorporates neighbor information via a cross-selective scan mechanism, where a neighbor-dependent system matrix is aggregated with distance-based weights and used to decode the system states within the target motion hypotheses. Finally, a regression head generates multiple socially plausible future trajectories, each with a corresponding probability. This dual-head prediction allows the model not only to forecast diverse possible futures but also to provide an estimation for each prediction. Extensive experiments on three datasets demonstrate that our method has the best average minADE20/minFDE20of 0.19/0.30 on ETH/UCY and the lowest ADE of 6.72 on SDD, surpassing existing baseline models. Our method also demonstrates superior computational efficiency, achieving the highest inference speed while maintaining the lowest FLOPs of 1.11G and a competitive parameter size of 0.12M. This optimal balance between prediction accuracy, computation, and memory efficiency ensures reliable real-time trajectory prediction in autonomous driving.