By Consentium IoT — 24 Jul 2025

Capturing Intelligence at the Edge: Building Image Datasets with EdgeVision and ESP32-CAM

In the world of edge AI and TinyML, one of the biggest roadblocks is collecting and labeling high-quality data — especially when you're working with constrained hardware like the ESP32-CAM. That's where EdgeVision, a library from Consentium IoT, steps in.

It transforms your humble ESP32-CAM board into a full-fledged, web-controlled image capture and labeling studio, making it remarkably easy to generate datasets right from the device — no PC, Raspberry Pi, or cloud connection required.

🔍 What Is EdgeVision?

EdgeVision is a modular library and browser-based tool that allows developers, students, and ML practitioners to:

Capture still images at fixed intervals via the ESP32-CAM
Label data with ZIP filenames in-browser
Bundle and download image datasets as ZIP archives
Use the data for training or real-time ML inference

Built for simplicity and speed, it turns the ESP32 into a mini data station for collecting real-world images — perfect for object detection, pose recognition, and environment monitoring.

🖼️ What the Interface Looks Like

Once the firmware is flashed and the device connected to your network, EdgeVision serves a live video stream and capture interface in the browser:

You can:

🎥 View the live camera feed
⏱️ Set a capture interval (e.g., every 1500 ms)
🏷️ Enter a ZIP filename as the label (e.g., "Class A", "Fall", "Walking")
🟢 Click Start to begin snapping images
🔴 Stop, 🔁 Reset, and ⬇️ Download when you're done

Each image is stored in memory and zipped — ready for machine learning workflows.

⚡ Getting Started in Minutes

🧩 Hardware Requirements

ESP32-CAM module (AI Thinker or similar)
Micro-USB programmer (with 5V and GND wired correctly)
USB cable
Optional: Battery or external 5V source for field capture

🔧 Installation via Arduino Library Manager

EdgeVision is officially available via the Arduino Library Manager, making it easy to install.

Open Arduino IDE
Go to Tools → Manage Libraries
Search for EdgeSense
Click Install

After installation, go to:
File → Examples → EdgeVision → CameraStreaming

Flash the sketch to your ESP32-CAM board. Open Serial Monitor to get the IP address and access the web interface.

🧠 Let’s Understand the Code

Let’s walk through the default CameraStreaming.ino sketch. It’s less than 30 lines of code, but does a lot behind the scenes.

1. Include the Library and Set Wi-Fi

#include <EdgeVision.h>

const char* ssid = "YOUR_WIFI_SSID";
const char* password = "YOUR_WIFI_PASSWORD";
EdgeVision cam;

This initializes the EdgeVision library and sets up your Wi-Fi credentials. The cam object controls everything: Wi-Fi, camera, and HTTP server.

2. Setup: Connect, Configure, Serve

void setup() {
  Serial.begin(115200);

  cam.initWiFi(ssid, password);
  cam.initCamera();
  cam.startCameraServer();

  Serial.println("Camera server started.");
}

cam.initWiFi(...): Connects your ESP32 to your local Wi-Fi network
cam.initCamera(): Initializes the OV2640 camera (resolution, buffers, pins)
cam.startCameraServer(): Starts an HTTP MJPEG server for live video + control

When this finishes, your ESP32 is ready to stream and capture from the browser.

3. Loop: Keep the Server Alive

void loop() {
  cam.keepServerAlive();
}

This non-blocking function ensures the internal web server continues handling requests and streaming without crashing.

🧠 Why EdgeVision Matters

Let’s be real: collecting clean, labeled data is a pain.

EdgeVision simplifies the most tedious part of your ML pipeline — especially when you're targeting embedded or offline devices.

Here’s why it matters:

1. 📊 Collect Real-World Data

Whether you're capturing gestures, motion patterns, object shapes, or lighting changes, collecting authentic, device-native data is critical. Training your model on what the sensor actually sees boosts real-world accuracy.

2. 🏷️ Label as You Capture

Instead of moving files, renaming folders, and batching images manually, EdgeVision labels the dataset automatically based on the ZIP name you provide. One ZIP = One class = Zero post-processing.

3. 🧠 Train Better TinyML Models

The tool integrates beautifully with EdgeModelKit — a Python library from Consentium IoT designed to preprocess, structure, and train ML models on the data captured by EdgeVision.

🌍 Use Cases in the Wild

🤸 Activity Classification: Standing, walking, falling — perfect for health monitoring
🧪 STEM Education: Let students build datasets and deploy models themselves
🌱 Smart Agriculture: Detect pests, rot, or leaf shapes in the field
🏗️ Industrial Monitoring: Log visual anomalies without sending full video
🐶 Pet Recognition: Teach your ESP32 to tell the difference between a cat and a dog

All of this, without an external computer in the loop.

⚙️ Under the Hood

EdgeVision uses:

The ESP32-CAM's OV2640 camera module for image capture
MJPEG streaming via a lightweight HTTP server
Timed snapshots with customizable interval control
SPIFFS or RAM to temporarily store and package images
JavaScript on the frontend to manage UI and ZIP creation

It’s optimized to run entirely offline on the local network, ideal for bandwidth-constrained environments.

🧰 Pro Tips

For better image quality, use natural lighting or add IR LEDs for night vision.
Capture images from multiple angles and lighting conditions to improve dataset diversity.
Use a tripod or mount to keep the camera fixed while collecting.
Run multiple ESP32-CAMs with different class labels on different devices to parallelize data capture.

📘 Learn More

📦 GitHub Repo: ConsentiumIoT/EdgeVision
🔬 Dataset Training Toolkit: EdgeModelKit
🧠 Tutorials & Docs: https://docs.consentiumiot.com

📫 Get in Touch

Questions or suggestions?
📬 Email: [email protected]
🌐 Website: https://consentiumiot.com

EdgeVision is open-source and MIT licensed. It’s part of a growing suite of edge AI tools developed by Consentium IoT Labs — powering smarter sensing, right where the data is born.