多模态多序列高效越野可行驶区域检测

Freespace detection plays a crucial role in autonomous trajectory prediction and path planning; however, the complexity of off-road environments and the irregularity of navigable boundaries limit improvements in detection accuracy, real-time performance, and generalization. To address these challenges, a multimodal multi-sequence efficient freespace detection network is proposed. This network leverages the strengths of Transformer and CNN, integrating a spatiotemporal adaptive gating unit to effectively fuse and enhance multimodal multi-sequence information from RGB images and LiDAR data. Additionally, an M-IDA module is incorporated to refine output features, further improving detection accuracy and generalization. Moreover, linear Transformer encoding and depthwise separable convolutions are employed to reduce computational complexity and achieve efficient inference. Four comparative experiments on the ORFD dataset demonstrate that, compared to baseline models, the proposed network achieves 2.3%, 1.2%, and 2% improvements in accuracy, F1-score, and IoU, respectively, while reducing inference time by 40.8%. Furthermore, ablation studies validate the effectiveness of each module, and additional evaluations on the ORFD test set, KITTI road dataset, and real-vehicle experiments further confirm the network’s generalization accuracy, efficiency and capability.