AI 中文总结
本文提出RaStream框架，结合雷达感知空间编码器与双状态因果时间优化模块，在边缘端实现毫米波雷达流人体网格恢复，在M4Human数据集上显著降低MVE，且延迟低至26.93 ms。
AI 中文摘要
毫米波（mmWave）雷达可为边缘应用提供隐私保护型人体感知能力，但在边缘设备上进行SMPL-X流恢复需要在轻量因果推理下实现精确的空间特征提取与时间稳定预测。稀疏的雷达反射信号使得密集网格恢复任务难度较大，而庞大的多尺度空间主干网络在处理体素化雷达张量时会产生较高计算成本，同时仍会因背景杂波弱化微弱的人体特征。逐帧网格估计会出现抖动问题，而通用时间模型往往会将缓慢变化的人体形态与快速的姿态、平移动态混淆。本文提出RaStream，这是一种可部署于边缘端的雷达张量流网格恢复框架，结合了雷达感知空间编码器与双状态因果时间优化模块。雷达感知空间结构（RaSS）编码器保留3D雷达结构，定位目标人体，提取以人体为中心的特征，并从短雷达窗口生成紧凑的雷达感知令牌。双状态时间模块将缓慢的形态状态与快速的运动状态分离：它通过令牌条件更新门累积形态特征以进行形状和性别估计，并通过因果循环状态跟踪动态运动。所得模型保持固定的流内存，避免全量缓冲。我们定义了时间采样参数（Tw, T, s），该参数可反映雷达观测密度、有限展开 horizon、预热/重放行为及输出速率权衡，并在M4Human数据集上评估重建精度、时间平滑度和边缘效率。RaSS-Base以更少参数将单窗口MVE从90.90 mm降至84.27 mm，优于RT-Mesh，而RaStream在随机分割协议下进一步将MVE降至72.05 mm。Jetson Orin Nano profiling显示，Base配置的FP32延迟为26.93 ms。
英文摘要
Millimeter-wave (mmWave) radar provides a privacy-preserving sensing modality for human motion analysis, yet continuous SMPL-X recovery on edge devices remains challenging because accurate reconstruction benefits from dense volumetric radar observations that preserve weak articulated reflections, while repeatedly processing such high-dimensional data over time incurs substantial computational cost. Temporal context is also necessary to resolve frame-local ambiguity, but SMPL-X attributes evolve at different rates, with body morphology remaining relatively stable and articulated pose, root orientation, and global translation changing rapidly. We present RaStream, a causal streaming framework that separates rich volumetric spatial perception from lightweight temporal reasoning. For each radar observation, a localization-conditioned spatial encoder exploits global scene context to estimate a coarse body location and focuses detailed reconstruction on a center-guided region of interest while preserving native 3D radar structure. Global and local features are then fused into a compact representation, allowing volumetric evidence to be extracted once per observation and temporal history to be propagated through compact representations and fixed-size persistent states. RaStream further maintains a slow morphology state for shape-related information and a fast motion state for pose, orientation, and translation. On the 661K-frame M4Human benchmark, RaStream reduces MVE from 90.90 mm for RT-Mesh to 84.27 mm with spatial-only inference and to 72.05 mm with causal temporal refinement. The Base configuration requires 26.93 ms per FP32 invocation on a Jetson Orin Nano. These results demonstrate practical edge-streaming radar mesh recovery through one-time volumetric evidence extraction and compact causal temporal propagation.
