FROM CAMERA TO DATASET

Every good model
starts in the field.

Capture a frame. Draw the people. Build your edge dataset.

Camera not connected

FRAMES COLLECTED0

READY FOR TRAINING0

PERSON BOXES0

TO LABEL0

ANNOTATION STUDIO

No frame selected

CAMERA SOURCE

CAPTURE SESSIONDATA SPLIT↑ Import

＋

Bring your camera into the picture.

Connect your ESP32-CAM or import a few images.
Your frames will appear here, ready to label.

Draw · move · resize

Drag to draw a person box. Drag a box to move it; drag its corners to resize. Shift+drag adds an overlapping box. Del removes selection · Ctrl Z undoes.

01 / COLLECT

Different sessions, better evidence. Capture validation and test images at different times or locations. Each session stays in one split to reduce leakage.

ONE CLASS: PERSON

SMALL MODEL, REAL BOUNDING BOXES

Train for the edge.

A compact, single-class detector with full INT8 export.

↓ Export dataset

EXPERIMENTAL BASELINE

ESP32 Person AABB

Grayscale depthwise CNN with a stride-8 detection head. It learns objectness, center offsets, width, and height, then applies non-maximum suppression.

BEFORE YOU TRAIN

Make the data count.

Draw a tight box around every visible person.
Include empty frames and mark them “No people.”
Use separate capture sessions for training, validation, and test.
Collect different distances, lighting, poses, backgrounds, and partial occlusions.

This is a trainable baseline, starting from random weights. It has not been benchmarked on your camera. For ESP32-S3, compare with Espressif’s ESPDet-Pico using the exported YOLO dataset.

One person center per output cell. Crowded scenes may need a stronger detector. Minimum to exercise the workflow: 3 train, 1 validation, 1 test image; useful evaluation needs many more.

Training runs

Best validation checkpoint · held-out test evaluation

EVIDENCE BEFORE ARCHITECTURE

A practical reading list.

ESP32 applications, true-box TinyML research, and maintained deployment tools.

↓ Research notes

“Person detection” can mean three different tasks.

Classification tells you if anyone is present. FOMO predicts object centers. AABB detection predicts the person’s position and extent. Results from ESP32-S3, P4, GAP9, or MAX78000 do not establish performance on the original ESP32-CAM.

CONNECT YOUR HARDWARE

One camera.
A continuous workflow.

Reuse your Container streamer, or pair a standalone capture camera.

AI-THINKER ESP32-CAM

ALREADY STREAMING TO CONTAINER?

Use your existing camera.

The lab can read your authenticated Container Cameras stream and save its current frame when you press Capture. Your existing firmware and recording continue to work. Access uses your Oaiflow session and Container's existing account permissions.

01 / Pair a standalone camera

Create a device token for your account. Pairing again invalidates the previous token. On a preflashed camera, connect to its UMFST-CAM setup Wi-Fi, open http://192.168.4.1/ on your phone, and enter your 2.4 GHz Wi-Fi details plus the LAB_DEVICE_TOKEN value. Settings use internal flash; no SD card is needed.

Download this configuration now; the token is shown only for this pairing.

02 / Flash the capture sketch

For an unflashed camera, download the sketch and place it in a folder named capture, alongside config.h. The config.h file is optional: without configured credentials, the camera starts its local Wi-Fi setup page.

↓ Download capture.ino

Select AI Thinker ESP32-CAM in Arduino IDE, enable PSRAM, upload over USB serial. Automatic-reset programming boards need no RST press or jumper. Leave RST released while writing; press and release it once only if a restart is needed after upload. The sketch uses the ESP32 Arduino core 3.x. JPEG capture falls back to native RGB565 for sensors without JPEG support.

03 / Capture from anywhere

Web Capture↓ queued commandESP32-CAM on Wi-Fi↓ JPEG or native RGB565 over HTTPSYour private labeling queue

Keep the board powered and connected. Its status turns green after the first poll. Press Capture frame to request a fresh frame.

Optional: a momentary button between GPIO13 and GND captures directly into button-session (training split). Do not use the SD card at the same time.

The supplied firmware collects images. Training runs on GMK. Running the exported detector on the camera is a separate integration step; see the deployment notes in the source.

Original ESP32 pin map included. An S3 camera needs its own pin mapping. No camera IP, port forwarding, or browser access to local HTTP is required.

Your field research workspace

Every good modelstarts in the field.

Bring your camera into the picture.

Train for the edge.

ESP32 Person AABB

Make the data count.

Training runs

A practical reading list.

“Person detection” can mean three different tasks.

One camera.A continuous workflow.

Use your existing camera.

01 / Pair a standalone camera

02 / Flash the capture sketch

03 / Capture from anywhere

Every good model
starts in the field.

One camera.
A continuous workflow.