HMambaOcc: Hierarchical Mamba for occupancy flow field prediction in autonomous driving under mixed traffic environments

Artikel i vetenskaplig tidskrift, 2026

Occupancy flow field prediction is essential for autonomous driving under mixed traffic environments, as it provides pivotal information for safe decision-making and planning. Although existing approaches show promise in forecasting surrounding agents, they suffer from limited computational efficiency and challenges in modeling complex interactions, making it difficult to fully capture the intrinsic relationships among various factors. To this end, we propose HMambaOcc, a hierarchical Mamba-based framework for occupancy and flow prediction. Specifically, to fully leverage available multimodal scene information, we introduce the dual-channel input module, incorporating both vectorized and visual channels. Then, the hierarchical Mamba encoder module is developed to efficiently extract latent features from dual-channel inputs with linear computational complexity, significantly improving computational efficiency. Next, we design the cascade interaction module based on the attention mechanism, aiming to promote the model's ability to extract both fine-grained local interaction and high-level global interaction across different input factors. Finally, the multi-branch decoder is presented to generate future occupancy and flow predictions. We conduct extensive experiments on the Waymo Open Motion Dataset to evaluate the effectiveness of our model. Quantitative results demonstrate that our framework exhibits competitive prediction accuracy with high computation efficiency against the state-of-the-art methods. Specifically, HMambaOcc achieves the highest score on AUCb (0.8164), ranks second on Soft-IoUb (0.4937) and Soft-IoUfgo (0.5506), and ranks third on Soft-IoUc (04001) and AUCfgo (0.7928), respectively. Moreover, it reduces inference time to 198.8ms, improving efficiency by 17.1 % and 8.8% over StrajNet and OFMPNet, respectively. Qualitative results demonstrate a consistent correspondence between occupancy prediction and flow prediction, further confirming that HMambaOcc produces effective and reasonable predictions. Additionally, HMambaOcc exhibits a decrease of only 6.38% and 5.11% in the primary metric AUCfgo when facing missing trajectories and unseen scenarios, respectively, demonstrating its robustness and generalization capability.