HDDPNet: Hierarchical Cascade and Dual-Stream Diffusion for Multi-Task Perception in Autonomous Driving
推荐理由
Environmental perception is fundamental to the safety and reliability of autonomous driving systems. Despite recent progress, existing frameworks often struggl…
核心判断
论文摘要（中文）
论文提出 HDDPNet 多任务自动驾驶感知网络。编码器使用 TransMamba 层级级联建模全局上下文，并以渐进式特征金字塔融合多尺度信息；解码器 ID3Net 用单步扩散式去噪和双流结构促进任务间协作。
Environmental perception is fundamental to the safety and reliability of autonomous driving systems. Despite recent progress, existing frameworks often struggle to achieve multi-granularity fusion of BEV spatial features in complex traffic scenarios. Furthermore, some approaches treat perception tasks separately and do not adequately model inter-task associations. To address these challenges, we propose HDDPNet, a multi-task perception model for autonomous driving, which adopts an encoder-decoder architecture to enhance semantic modeling and cross-task collaboration. In the encoding stage, we introduce a Hierarchical Cascade Encoder (HCE), which incorporates a TransMamba module to enhance global context modeling and feature representation. In addition, a Progressive Feature Pyramid (PFP) module is designed to facilitate multi-scale and fine-grained information fusion. In the decoding stage, we develop an Isomorphic Dual-stream Diffusion Decoder Network (ID3Net), which adopts a diffusion-inspired single-step denoising process to improve feature robustness and promote cross-task collaboration by sharing semantic and geometric information within a dual-stream architecture. Experimental results show that the proposed method achieves significant improvements in multiple perception tasks: NDS is increased by 10.09% in 3D object detection; AMOTA is improved by 7.29% in multi-object tracking; and mAP is increased by 2.17% in online mapping. These results demonstrate that HDDPNet consistently improves multi-task perception performance, providing a reliable solution for autonomous driving.
研究动机（中文总结）
复杂交通场景下，现有方法难以同时融合 BEV 特征的不同空间粒度；把检测、跟踪、地图等任务分开处理，也未充分利用任务之间的关联。
To address these challenges, we propose HDDPNet, a multi-task perception model for autonomous driving, which adopts an encoder-decoder architecture to enhance semantic modeling and cross-task collaboration.
创新与贡献（中文总结）
把层级上下文建模、多尺度特征融合与扩散式双流解码结合到统一的多任务感知架构，显式传递语义和几何信息。
To address these challenges, we propose HDDPNet, a multi-task perception model for autonomous driving, which adopts an encoder-decoder architecture to enhance semantic modeling and cross-task collaboration.
方法与证据
方法（中文总结）
编码阶段由 HCE/TransMamba 提取全局上下文，并由 PFP 融合多尺度特征；解码阶段 ID3Net 以单步去噪生成稳健表示，通过双流共享语义与几何信息。
To address these challenges, we propose HDDPNet, a multi-task perception model for autonomous driving, which adopts an encoder-decoder architecture to enhance semantic modeling and cross-task collaboration.
实验结果（中文总结）
摘要报告 3D 检测 NDS 提升 10.09%，多目标跟踪 AMOTA 提升 7.29%，在线建图 mAP 提升 2.17%；这些是摘要给出的相对提升，具体实验设置仍应结合论文表格核对。
Experimental results show that the proposed method achieves significant improvements in multiple perception tasks: NDS is increased by 10.09% in 3D object detection; AMOTA is improved by 7.29% in multi-object tracking; and mAP is increased by 2.17% in online mapping.