不装摄像头,Wi-Fi 也能”看见”你?我来拆解这个 30 美元的实验

2026年11月14日 · AOAI AI

No Camera Needed? Let Me Break Down This $30 Wi-Fi Experiment

今天想和大家聊一个最近在开源圈很火的项目:RuView。它的卖点听上去像监控电影——把电脑摄像头完全挡住,只靠家里的 Wi-Fi 信号,就能判断一个人是在深蹲、坐着还是走动。我看到的一次公开实测里,整体准确率约为 79.5%。我想做的不是跟着喊”黑科技”,而是把原理、实测数据和它的水分一起讲清楚。先说明一点:文中的数字来自那次公开实测,我自己没有复现,所以请当作参考,不要当作我的测试结论。

Today I want to talk about a project that has been all over the open-source scene lately: RuView. Its pitch sounds like a surveillance thriller: cover the computer’s webcam completely, and from nothing but home Wi-Fi signals, tell whether someone is squatting, sitting or walking. In one public test I came across, overall accuracy was about 79.5%. I do not want to just shout “magic tech”; I want to lay out the principle, the test data and the hype together. One note first: the numbers here come from that public test, and I have not reproduced them myself, so treat them as reference, not as my own test conclusions.


先讲原理,其实不复杂。路由器发出的信号不是走直线的,它会在墙壁、家具和人体之间不断反射,并分布在几十个子频段上。接收端为了正确解码数据,会计算每个子频段的衰减和相位偏移,这组数据叫信道状态信息,简称 CSI。你一动,电磁波的反射路径就变了,CSI 的数值也会出现有规律的波动。要让 AI 读懂这些波动,需要一位”老师”:这个思路来自卡内基梅隆大学 2022 年的研究,先用摄像头跑姿态识别给动作打标签,同时记录 Wi-Fi 的 CSI,训练模型只凭波形去预测姿态,训练完就能把摄像头撤掉。

First, the principle, which is not complicated. A router’s signal does not travel in a straight line; it bounces among walls, furniture and bodies, spread across dozens of sub-bands. To decode data correctly, the receiver computes each sub-band’s attenuation and phase shift, and that set of numbers is channel state information, or CSI. When you move, the reflection paths change and CSI fluctuates in recognizable patterns. To make an AI read those fluctuations you need a “teacher”: the idea comes from a 2022 Carnegie Mellon study, where a camera running pose estimation labels the movements while Wi-Fi CSI is recorded, a model learns to predict poses from waveforms alone, and afterwards the camera can be taken away.


这次实测最值得讲的,是它先打了假。实测者下载了项目的预训练模型,放到自己的房间里,发现在没见过的新环境里准确率只有 11%。而网上疯传的”Wi-Fi 实时生成人体 3D 骨架”演示,骨架其实来自摄像头,Wi-Fi 的部分只是模拟效果。所以我的第一个提醒是:看开源项目的演示视频,要多问一句”这是真实输出,还是效果图”。对 CSI 这类对环境极度敏感的技术,换个房间就等于换了一套数据,宣传里的数字很难直接套用。

The most worth-telling part of this test is that it debunked first. The tester downloaded the project’s pretrained model and ran it in his own room, where, in an environment it had never seen, accuracy was only 11%. In the viral demo of “Wi-Fi generating a live 3D human skeleton,” the skeleton actually came from a camera and the Wi-Fi part was simulated. So my first reminder: when you watch an open-source demo, ask one more question, whether it is real output or a rendering. For a technology as environment-sensitive as CSI, a new room is effectively a new dataset, and promotional numbers rarely transfer directly.


接着是低成本的复现方案。硬件是三块约 10 美元的 ESP32-S3 开发板,合计大约 30 美元,刷入修改过的 RuView 固件,让板子之间也能互相监听,在客厅里织出 9 条交叉的感知线。训练在浏览器里完成:用 Google MediaPipe 识别动作当”老师”,同时记录三块板子的 Wi-Fi 信号,采样训练大约 9 分钟。这套流程的门槛之低,是我觉得最有冲击力的地方,一个普通爱好者就有机会做同类实验。

Next, the low-cost rebuild. The hardware is three ESP32-S3 boards at about $10 each, roughly $30 in total, flashed with modified RuView firmware so the boards can also listen to each other, weaving nine crossing sensing lines across the living room. Training happens in the browser: Google MediaPipe recognizes the movements as the “teacher” while the three boards log Wi-Fi signals, for about nine minutes of sampling and training. The low barrier of this pipeline is what I find most striking; an ordinary hobbyist could run a similar experiment.


遮住摄像头后的盲测,整体准确率是 79.5%。分项来看:空房间检测 100%,判断屋里有没有人 96%,跳跃 97%,坐姿 93%,原地旋转 92%,站立 84%,深蹲 79%,走动 70%。短板是挥手,只有 13%,经常被误判成跳跃,原因是手臂的截面积太小,扰动不了多少信号。规律不难理解:动作幅度越大、挡住的信号越多,越容易识别。所以看整体数字时一定要看分项,79.5% 不代表每一种动作都是这个水平。

In the blind test with the camera covered, overall accuracy was 79.5%. By category: empty room 100%, presence detection 96%, jumping 97%, sitting 93%, spinning in place 92%, standing 84%, squatting 79% and walking 70%. The weak spot is waving at just 13%, often mistaken for jumping, because an arm’s cross-section is too small to disturb much signal. The pattern is easy to grasp: the bigger the motion and the more signal blocked, the easier the recognition. So always read the category numbers; 79.5% does not mean every action performs at that level.


再说大家最关心的隐私。Wi-Fi 信号可以穿墙,理论上有人在墙外放几块廉价开发板,监听路由器的广播,就能推断屋里有没有人,甚至摸清作息。专门针对 Wi-Fi 感知的 IEEE 802.11bf 标准已经发布,将来家用路由器可能原生具备这类能力。但也要看到好的一面:它是不需要摄像头的无感传感器,适合室内安防、独居老人的日常监测和跌倒检测,RuView 也内置了跌倒检测。需要说明的是,”墙外监听”在这次实测里是理论推断,并没有现场演示。

Now the thing everyone cares about, privacy. Wi-Fi passes through walls, so in theory someone could put a few cheap boards outside, listen to the router’s broadcasts, and infer whether anyone is home and even work out routines. The IEEE 802.11bf standard, aimed specifically at Wi-Fi sensing, has already been published, so home routers may eventually have such abilities built in. But the upside is real too: it is a camera-free, unobtrusive sensor that suits indoor security, daily monitoring for elderly people living alone, and fall detection, which RuView already includes. Note that eavesdropping from outside the wall is a theoretical inference in this test, not something demonstrated live.


最后说我的看法。这个模型对环境高度敏感,挪动开发板的位置、房间里多进来一个人,都需要重新采集数据训练,离”开箱即用的通用监控”还有距离。所以对普通人来说,它现在更像一个有趣的实验和一个值得关注的趋势,而不是迫在眉睫的威胁。我觉得真正值得学的是方法:用摄像头当老师,训练一个以后不再需要摄像头的传感器,成本低、思路通用。如果你想动手,请只在自己家里、对自己的设备做,并遵守当地法律,不要未经同意监测他人。最后问大家一句:你敢在家里开一个”Wi-Fi 雷达”吗?欢迎在评论区聊聊。

Finally, my view. The model is highly environment-sensitive: move a board or let another person into the room and you need to collect and train on new data, which is a long way from out-of-the-box general surveillance. For most people it is, for now, an interesting experiment and a trend worth watching rather than an imminent threat. What I think is really worth learning is the method: using a camera as the teacher to train a sensor that no longer needs one is cheap and widely applicable. If you want to try it, do it only in your own home on your own equipment, follow local law, and do not monitor others without consent. One last question for you: would you dare to turn on a “Wi-Fi radar” at home? Let me know in the comments.