9 美元的 ESP32 能看穿墙壁?拆解 4.8 万 Star 的 WiFi 感知平台 RuView$9 ESP32 Can See Through Walls? Teardown of 48,000-Star WiFi Sensing Platform RuView


AI free token/AI免费积分

项目卡片

  • 项目:RuView[1]
  • 状态:v0.6.2-esp32 / 47,993 Stars / MIT / 活跃开发中
  • 一句话判断:WiFi 感知领域当前最完整的开源方案,工程深度远超同类,但多项高级功能仍处于早期实验阶段。

你的路由器每时每刻都在发射无线电波。当一个人走进房间,哪怕隔着墙壁,电波也会被他的身体散射、反射、吸收——这些变化是可测量的。

RuView 用 9 美元一块的 ESP32-S3 捕获 WiFi 信道状态信息(CSI),通过信号处理到神经网络的完整流水线,把电波扰动还原成有用信息:有人吗、呼吸频率多少、心率多少。

我翻完仓库后的第一反应是工程深度远超预期——15 个 Rust crate、1.2 万行 C 固件、79 篇 ADR,在开源 WiFi 感知领域几乎是独一无二的体量。

没有摄像头,没有穿戴设备,不联网。核心成本就是一块 ESP32。

RuView 实时 WiFi CSI 信号生成的 3D 骨架可视化

它不是什么

在展开之前先说清楚 RuView 不是什么。

它不是读心术。WiFi CSI 感知的本质是多径效应分析——人的身体改变了电磁波传播路径,ESP32 捕获这些变化并提取模式。精度受物理条件限制:单个 AP 大约区分 3-5 人,穿墙感知有效距离约 5 米。

它也不是即插即用的消费级产品。当前标注 Beta,最近一次更新还修复了 Timer Service 栈溢出导致启动 1 秒后崩溃的问题,部分 ADR 还是 Proposed 状态。

它是一个工程原型——但作为一个开源工程原型,它的完成度令人惊讶。

架构:三层分离

RuView 的系统分三层,每一层都独立可运行:

底层是 ESP32 固件(C 语言,约 1.2 万行)。从 WiFi 芯片提取 CSI 原始数据,做边缘信号处理(呼吸/心跳检测、跌倒判定),通过 UDP 发送数据帧。最新版引入了 ADR-081 五层架构(Radio Abstraction Layer / Adaptive Controller / Mesh Sensing Plane / On-device Feature Extraction / Rust Handoff),让同一套控制器适配不同射频芯片。60 字节紧凑特征包(rv_feature_state_t)把带宽从 ~100 KB/s 压到 300 B/s,压缩比 99.7%。

中间层是 Rust 信号处理引擎(15 个 crate,约 15.8 万行)。wifi-densepose-signal 实现 Hampel 滤波、相位解卷绕、Fresnel 几何、微多普勒提取;wifi-densepose-ruvector 集成子载波选择、注意力机制、Stoer-Wagner 最小割人员分离;wifi-densepose-train 提供从数据收集到模型训练的完整流水线。

上层是可视化与 API 服务。Rust Axum 服务器 + Three.js 前端,Docker 一键启动,不需要硬件就能用模拟数据体验。

RuView 的 Web 管理界面,实时显示感知数据

核心能力与成熟度

RuView 声称支持的能力非常多,从存在检测到姿态估计到穿墙搜救,跨度极大。我翻了一遍源码和测试,按实际验证情况把它们分成三个档位。

已验证可用

存在检测:60,630 个过夜样本,100% 准确率。推理 0.012 ms,单台 Mac Mini 处理 1,600+ 节点。预训练模型已发 HuggingFace,4-bit 量化仅 8 KB。

生命体征:呼吸 6-30 BPM、心率 40-120 BPM,基于 FFT 频谱分析。支持 60 GHz 毫米波雷达融合,实测 52 cm 距离心率 75 bpm。

跌倒检测:角速度阈值 15.0 rad/s²(正常行走不触发),3 帧去抖 + 5 秒冷却,实测 0 误报。

快速上手:Docker 方案确实做到了一条命令启动。docker pull ruvnet/wifi-densepose:latest && docker run -p 3000:3000 ruvnet/wifi-densepose:latest,打开浏览器即可看到模拟数据驱动的完整界面。

基本可用但有限制

姿态估计:WiFlow 架构(TCN + axial attention),17 个 COCO 关键点。无监督训练下精度有限;配合摄像头做监督训练可达 92.9% PCK@20(需 5 分钟收集配对数据)。无摄像头方案精度会打折。

多人追踪:Stoer-Wagner 最小割算法修复了长期 bug(Issue #348),多人融合用 GATv2 注意力机制,但效果高度依赖房间布局和节点部署。

多频率 mesh:ESP32 在 6 个 WiFi 信道间跳频,甚至利用邻居路由器做被动雷达。功能代码存在,但部署复杂度显著上升。

仍在纸面上

穿墙搜救(WiFi-MAT):代码实现了多 AP 三角定位和伤情分类,但"穿透 30cm 混凝土"的能力没有实测数据支撑。

65 个边缘 WASM 模块:代码存在、609 个测试通过,但从手势识别到情感检测,很多更像是概念验证,缺少真实场景测试报告。

跨环境泛化(MERIDIAN):ADR-027 描述了对抗性领域泛化方案(1,858 行代码、72 个测试),但缺少跨房间/跨建筑的实测验证。

信号处理管线的真实设计

抛开那些宏大的应用场景,我认为 RuView 最有价值的部分是它的信号处理管线。从源码中提炼出的真实数据流如下:

从 WiFi 信号到人体感知的数据流:ESP32 捕获 CSI → Rust 信号处理管线 → 姿态/生命体征/存在检测输出

WiFi 路由器 → ESP32-S3 mesh 捕获 CSI(信道 1/6/11,TDM 协议)
    ↓ 多频段融合(3 信道 × 56 子载波 = 168 虚拟子载波)
    ↓ 多基站融合(N×(N-1) 链路 → 注意力加权跨视点嵌入)
    ↓ 一致性门控(Z-score 相位一致性 + 滞后门控)
    ↓ 信号处理(Hampel / Fresnel / BVP / 频谱图)
    ↓ AI 骨干网(RuVector:注意力 + 图算法 + 场模型)
    → 输出:姿态 + 呼吸 + 心率 + 房间指纹

几个值得注意的工程决策:

  • CSI 而非 RSSI:大多数 WiFi 感知方案只利用 RSSI,精度有限。RuView 直接提取完整 CSI——每个子载波的幅度和相位——信息量大几个数量级。代价是只支持 ESP32-S3 和研究级 NIC,不支持 ESP32-C3 和原版 ESP32。

  • 一致性门控:不是所有时刻的 CSI 都可信。coherence gate 用 Z-score 相位一致性评分加滞后门控,自动决定接受、预测或拒绝每帧数据,这是系统能稳定运行数天无需调参的关键。

  • Adaptive Controller(ADR-081):三层闭环控制——快环(200ms)调采样频率、中环(1s)切信道、慢环(~30s)重校准基线。真正的控制系统设计,不是 if-else。

开发体验

如果你想上手试一下,几个关键信息:

最低门槛:Docker 方案,零硬件。拉镜像运行,浏览器打开 localhost:3000 即可体验完整 UI(模拟数据)。

入门硬件:一块 ESP32-S3(8MB Flash),约 9 美元。串口烧录固件,用 provision.py 配 WiFi,固件自动通过 UDP 发送 CSI 数据。

推荐配置:2-3 块 ESP32-S3 组 mesh,连接同一路由器,多节点显著提升检测精度。

测试覆盖是强项:1,463 个测试通过。Rust 侧 1,031+,Python 侧有确定性验证脚本(参考信号 + SHA-256 哈希),固件侧有 QEMU 模拟和主机端单元测试。

值得警惕的地方

这个项目有几个需要注意的问题:

scope 膨胀。不到一年半时间,项目从"WiFi 姿态估计"膨胀到涵盖睡眠监测、灾害搜救、材料分类、被动雷达、情感检测等几十个方向。79 篇 ADR 说明文档化做得好,但方向太多,任何一个的深度都可能受限。

CLAUDE.md 暴露的信息。项目的 CLAUDE.md 主要描述 Claude Flow 多智能体编排框架,而非 WiFi 感知本身——说明开发大量依赖 AI 辅助。15.8 万行 Rust 中有多少经过真实场景验证,需要你自己判断。

安全修复记录。CHANGELOG 中记录了 SQL 注入、XSS、命令注入、路径穿越等安全修复。对于一个处理敏感感知数据(生命体征、人员位置)的系统,这些修复说明早期安全实践有待加强。

单作者项目的可持续性。仓库 owner rUv 是项目的主要贡献者。这个规模的代码库(15.8 万行 Rust + 1.2 万行 C + Python + JS),长期维护是一个现实挑战。

总结

RuView 是 WiFi 感知开源领域目前最完整的实现。存在检测和生命体征提取已经具备实用价值,如果你需要在隐私敏感场景做人员感知,这是一个值得认真评估的方案。

项目当前真正能打的是基础感知能力和工程架构。穿墙搜救、情感检测、量子启发计算这些方向展示了技术天花板,但距离落地还有相当距离。建议从存在检测和生命体征开始验证——代码和测试最扎实。


如果这篇对你有用,建议点个关注。我会持续把 GitHub 上值得用的 AI 工具拆成「最短上手闭环 + 坑点清单 + 可复用配置」,让你少走弯路。

引用链接

[1]RuView: https://github.com/ruvnet/RuView

 

Project Card

Project: RuView [1]

Status: v0.6.2-esp32 / 47,993 Stars / MIT / Actively Developed

One-Sentence Verdict: The most complete open-source solution in WiFi sensing today, with engineering depth far exceeding similar projects, though several advanced features remain in early experimental stages.

Your router emits radio waves every moment. When a person enters a room, even through walls, those waves scatter, reflect, and are absorbed by their body—these changes are measurable.

RuView uses a $9 ESP32-S3 board to capture WiFi Channel State Information (CSI), then runs a complete pipeline from signal processing to neural networks to convert radio wave disturbances into useful information: is someone there? What's the breathing rate? What's the heart rate?

My first reaction after reviewing the repository was that the engineering depth far exceeded expectations—15 Rust crates, 12,000 lines of C firmware, and 79 ADRs, a scale almost unique in the open-source WiFi sensing field.

No cameras, no wearable devices, no internet connection. The core cost is a single ESP32 board.

RuView 实时 WiFi CSI 信号生成的 3D 骨架可视化

What It Is Not

Before diving in, let's clarify what RuView is not.

It's not mind-reading. WiFi CSI sensing is fundamentally multipath effect analysis—the human body alters electromagnetic wave propagation paths, and the ESP32 captures these changes to extract patterns. Accuracy is physically constrained: a single AP can distinguish roughly 3-5 people, and wall-penetration sensing has an effective range of about 5 meters.

It's also not a plug-and-play consumer product. Currently marked as Beta, the latest update even fixed a Timer Service stack overflow that caused crashes 1 second after boot, and some ADRs are still in Proposed status.

It is an engineering prototype—but as an open-source engineering prototype, its level of completion is impressive.

Architecture: Three-Layer Separation

RuView's system is divided into three layers, each independently runnable:

The bottom layer is the ESP32 firmware (C language, ~12,000 lines). It extracts raw CSI data from the WiFi chip, performs edge signal processing (breathing/heartbeat detection, fall判定), and sends data frames via UDP. The latest version introduced the ADR-081 five-layer architecture (Radio Abstraction Layer / Adaptive Controller / Mesh Sensing Plane / On-device Feature Extraction / Rust Handoff), enabling the same controller to adapt to different RF chips. The 60-byte compact feature package (rv_feature_state_t) compresses bandwidth from ~100 KB/s to 300 B/s—a 99.7% compression ratio.

The middle layer is the Rust signal processing engine (15 crates, ~158,000 lines). wifi-densepose-signal implements Hampel filtering, phase unwrapping, Fresnel geometry, and micro-Doppler extraction; wifi-densepose-ruvector integrates subcarrier selection, attention mechanisms, and Stoer-Wagner minimum cut for multi-person separation; wifi-densepose-train provides a complete pipeline from data collection to model training.

The top layer is visualization and API services. A Rust Axum server + Three.js frontend, startable with one Docker command, allowing you to experience the full interface with simulated data without any hardware.

RuView 的 Web 管理界面,实时显示感知数据

Core Capabilities and Maturity

RuView claims to support a wide range of capabilities, from presence detection to pose estimation to wall-penetration search and rescue. After reviewing the source code and tests, I've categorized them into three tiers based on actual verification status.

Verified and Working

Presence Detection: 60,630 overnight samples, 100% accuracy. Inference takes 0.012 ms, and a single Mac Mini can handle 1,600+ nodes. Pre-trained models are published on HuggingFace, with 4-bit quantization at only 8 KB.

Vital Signs: Breathing 6-30 BPM, heart rate 40-120 BPM, based on FFT spectral analysis. Supports 60 GHz mmWave radar fusion, with实测 heart rate of 75 bpm at a 52 cm distance.

Fall Detection: Angular velocity threshold of 15.0 rad/s² (normal walking doesn't trigger), with 3-frame debouncing + 5-second cooldown, achieving 0 false positives in testing.

Quick Start: The Docker solution truly achieves one-command startup. docker pull ruvnet/wifi-densepose:latest && docker run -p 3000:3000 ruvnet/wifi-densepose:latest , then open a browser to see the complete UI driven by simulated data.

Basically Usable but with Limitations

Pose Estimation: WiFlow architecture (TCN + axial attention) with 17 COCO keypoints. Limited accuracy under unsupervised training; with camera-assisted supervised training, it can reach 92.9% PCK@20 (requires 5 minutes to collect paired data). Without cameras, accuracy drops.

Multi-Person Tracking: The Stoer-Wagner minimum cut algorithm fixed a long-standing bug (Issue #348), and multi-person fusion uses GATv2 attention mechanisms, but results heavily depend on room layout and node deployment.

Multi-Frequency Mesh: ESP32s hop between 6 WiFi channels, even using neighboring routers as passive radar. The feature code exists, but deployment complexity increases significantly.

Still on Paper

Wall-Penetration Search and Rescue (WiFi-MAT): Code implements multi-AP trilateration and injury classification, but the "penetration through 30cm concrete" capability lacks实测 data support.

65 Edge WASM Modules: Code exists and 609 tests pass, but from gesture recognition to emotion detection, many seem more like proof-of-concept, lacking real-world testing reports.

Cross-Environment Generalization (MERIDIAN): ADR-027 describes an adversarial domain generalization approach (1,858 lines of code, 72 tests), but lacks cross-room/cross-building validation.

The Real Signal Processing Pipeline

Setting aside those grand application scenarios, I believe RuView's most valuable part is its signal processing pipeline. The actual data flow distilled from the source code is as follows:

从 WiFi 信号到人体感知的数据流:ESP32 捕获 CSI → Rust 信号处理管线 → 姿态/生命体征/存在检测输出

Several noteworthy engineering decisions:

CSI rather than RSSI: Most WiFi sensing solutions only use RSSI, which has limited accuracy. RuView directly extracts complete CSI—the amplitude and phase of each subcarrier—with information volume orders of magnitude greater. The cost is that it only supports ESP32-S3 and research-grade NICs, not ESP32-C3 or the original ESP32.

Coherence Gating: Not all CSI readings are trustworthy at all times. The coherence gate uses Z-score phase consistency scoring with hysteresis gating to automatically decide whether to accept, predict, or reject each frame of data—this is key to the system running stably for days without parameter tuning.

Adaptive Controller (ADR-081): Three-layer closed-loop control—fast loop (200ms) adjusts sampling frequency, medium loop (1s) switches channels, slow loop (~30s) recalibrates baseline. This is real control system design, not if-else logic.

Development Experience

If you want to try it out, here are several key pieces of information:

Lowest barrier to entry: Docker solution, zero hardware. Pull the image and run it, open browser to localhost:3000 to experience the complete UI (simulated data).

Entry-level hardware: A single ESP32-S3 (8MB Flash), approximately $9. Serial flash the firmware, configure WiFi with provision.py, and the firmware will automatically send CSI data via UDP.

Recommended configuration: 2-3 ESP32-S3 boards in a mesh, connected to the same router. Multi-node setups significantly improve detection accuracy.

Test coverage is a strength: 1,463 tests pass. 1,031+ on the Rust side, Python side has deterministic verification scripts (reference signals + SHA-256 hashes), and the firmware side has QEMU simulation and host-side unit tests.

Things to Watch Out For

This project has several issues to be aware of:

Scope creep: In less than a year and a half, the project expanded from "WiFi pose estimation" to covering dozens of directions including sleep monitoring, disaster rescue, material classification, passive radar, and emotion detection. 79 ADRs show good documentation practices, but with so many directions, depth in any single area may be limited.

CLAUDE.md reveals information: The project's CLAUDE.md primarily describes the Claude Flow multi-agent orchestration framework, not WiFi sensing itself—indicating development relies heavily on AI assistance. You'll need to judge for yourself how much of the 158,000 lines of Rust code has been validated in real scenarios.

Security fix history: CHANGELOG records fixes for SQL injection, XSS, command injection, path traversal, and other vulnerabilities. For a system handling sensitive sensing data (vital signs, personnel locations), these fixes indicate early security practices need strengthening.

Single-author project sustainability: Repository owner rUv is the primary contributor. Maintaining a codebase of this scale (158,000 lines of Rust + 12,000 lines of C + Python + JS) long-term is a real challenge.

Summary

RuView is currently the most complete implementation in the open-source WiFi sensing field. Presence detection and vital signs extraction already have practical value—if you need personnel sensing in privacy-sensitive scenarios, this is a solution worth serious evaluation.

The project's真正 competitive strengths lie in its basic sensing capabilities and engineering architecture. Directions like wall-penetration rescue, emotion detection, and quantum-inspired computing demonstrate the technical ceiling, but are still far from落地. I recommend starting with presence detection and vital signs—where the code and tests are most solid.

If this was helpful, consider following. I'll continue breaking down useful GitHub AI tools into "shortest setup loop + pitfall checklist + reusable configurations" to help you avoid detours.

Reference Links

[1] RuView: https://github.com/ruvnet/RuView

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry