Using two ESP32-S3 modules, real-time human detection and blinking

[QuHuiWan]

https://github.com/user-attachments/assets/0d091e57-c1fb-4eef-885c-1779981a83df
As shown in the video, I used two ESP32-S3 modules, but the real-time human body detection flickers

[QuHuiWan]

I configured two modules and obtained RSSI and AMP data through a serial port tool

python firmware/esp32-csi-node/provision.py --port COM15 --ssid "TP-LINK_D05A" --password "88888888" --target-ip 192.168.1.100 --target-port 5005 --tdm-slot 0 --tdm-total 2

[executionreverted]

i have same issue. i put 4 esp32, they work correctly, but i have flicker

i tried training section to record csi and train but it doesnt do anything. it doesnt start training

[xiaoji525]

我也有同样的问题。我插了4个ESP32，它们都能正常工作，但是屏幕会闪烁。

我尝试在训练部分录制 CSI 并进行训练，但没有任何反应，训练根本没有开始。
请问人的数量是正确的吗

[shushu789]

I encountered the same problem using three ESP32-S3 chips; the character skeletons were jumping around erratically and were completely unstable, making me even suspect that these character skeletons were virtual data.

20260305_182816.mp4

[ruvnet]

How did you calibrate your esp32s and what settings are you using? Please share them. Sounds like you didn’t do any room specific calibration.

[ruvnet]

Thanks for reporting this, and thanks to everyone sharing their experiences. The flickering/jumping you're seeing is expected behavior before room-specific calibration — here's why and how to fix it.

Why It Flickers

ESP32 CSI signals are extremely environment-dependent. Raw CSI amplitude varies with:

• Room geometry, furniture, wall materials
• ESP32 placement (height, orientation, distance between nodes)
• WiFi channel noise, interference from other APs
• Frame-to-frame jitter inherent in ~10 FPS CSI sampling

The default threshold-based classifier uses static values that work for a "generic room" but will produce noisy, unstable classifications in your specific environment. This is what causes the skeleton to jump between states.

How to Fix: Adaptive Classifier (ADR-048)

We recently merged the Adaptive CSI Activity Classifier (PR #139) which addresses exactly this problem with a 3-stage smoothing pipeline:

Step 1: Record Training Data

Record ~30 seconds of each activity state in your room:

# Empty room (leave the room!)
curl -X POST http://localhost:3000/api/v1/recording/start -H 'Content-Type: application/json' -d '{"id":"train_empty_room"}'
# Wait 30s, then:
curl -X POST http://localhost:3000/api/v1/recording/stop

# Standing still in the room
curl -X POST http://localhost:3000/api/v1/recording/start -H 'Content-Type: application/json' -d '{"id":"train_still_presence"}'
# Wait 30s, then:
curl -X POST http://localhost:3000/api/v1/recording/stop

# Walking around
curl -X POST http://localhost:3000/api/v1/recording/start -H 'Content-Type: application/json' -d '{"id":"train_walking"}'
# Wait 30s, then:
curl -X POST http://localhost:3000/api/v1/recording/stop

# Active movement (waving arms, exercising)
curl -X POST http://localhost:3000/api/v1/recording/start -H 'Content-Type: application/json' -d '{"id":"train_active_movement"}'
# Wait 30s, then:
curl -X POST http://localhost:3000/api/v1/recording/stop

Step 2: Train the Classifier

curl -X POST http://localhost:3000/api/v1/adaptive/train

This trains a multiclass logistic regression on 15 features extracted from your room's specific CSI patterns. Training takes <1 second for 3,000+ frames.

Step 3: Verify

curl http://localhost:3000/api/v1/adaptive/status

You should see per-class accuracy and the model will auto-load on server restart.

Multi-ESP32 Setup Tips

For 2-4 node setups:

1. Stagger TDM slots — make sure each ESP32 has a unique --tdm-slot (0, 1, 2, 3) with --tdm-total matching the count
2. Consistent placement — keep nodes at the same height (~1.2m recommended), spaced around the room perimeter
3. Same WiFi channel — all nodes should connect to the same AP on the same channel
4. Re-train after moving — if you move any ESP32, re-record and re-train

Re: Training Not Starting (@executionreverted)

The training endpoint requires recordings with filenames matching train_* pattern. Make sure:

1. The recording ID starts with train_ (e.g., train_empty, train_still, train_walking, train_active)
2. You have at least 2 different activity classes recorded
3. The recordings are in data/recordings/ directory
4. Check server logs for errors: the training endpoint returns a JSON response with accuracy metrics or error details

If training still doesn't start, please share the server log output and the response from the /api/v1/adaptive/train endpoint.

Built-in Smoothing

Even without adaptive training, the latest version includes:

• Adaptive baseline subtraction (EMA α=0.003, ~30s time constant)
• Hysteresis debounce (4 consecutive frames needed to change state)
• Vital signs smoothing (21-frame median + trimmed mean + dead-band filtering)

Make sure you're on the latest main branch to get these improvements.

[QuHuiWan]

I indeed haven't calibrated it. I will attempt to calibrate it using the steps you provided and then provide feedback

[QuHuiWan]

Additionally, do I need a clean room? Can I place items such as tables and vases in it

[1xhh1]

Is this project written by Claude AI?

[dlopez-create]

Hey @ruvnet could you confirm that it is possible to really have a stable train set to see a decent render?

[ruvnet]

You'll need to calibrate this based on your environment. I've only optimized it for me.

This is just a random project, not a job or primary focus for me.

[hzcoolwind]

⚠️ 欺诈警报：技术审计显示该项目是一个无效的人工智能生成假象，使用假 CSI 数据，没有训练模型，且怀疑存在星象膨胀
https://github.com/deletexiumu/wifi-densepose

[ba-mao]

It's been stuck in the running state for over half an hour.

[tOMC5188]

What esp32-s3 modules are people using? big dev-kits, small standalone, battery-powered?