EdgeBlink

EdgeBlink is a real-time object detection system running entirely on the ESP32-CAM using a quantized TensorFlow Lite model. Optimized with the EON™ Compiler, it delivers ~1.4 FPS with minimal memory usage, displays results on an OLED, and supports both USB and standalone battery-powered operation for portable AI-IoT applications.

sorry abt the intensity of the power led...and poor photography

🚀 Features

• Real-time Object Detection: Detects bitter gourd, onions, and tomatoes
• High Accuracy: 100% F1 score on validation set
• Optimized Performance: 695ms inference time with quantized int8 model
• Low Memory Footprint: 239.4K RAM usage, 81.1K flash usage
• OLED Display: Real-time detection results visualization
• Independent Power Supply: MP1584 buck converter support for standalone operation

📊 Model Performance

Training Results

• F1 Score: 100.0%
• Precision: 1.00
• Recall: 1.00
• Model Version: Quantized (int8)

Confusion Matrix

The model achieves perfect classification across all three classes:

Class	Background	Bitter Gourd	Onion	Tomato
Background	100%	0%	0%	0%
Bitter Gourd	0%	100%	0%	0%
Onion	0%	0%	100%	0%
Tomato	0%	0%	0%	100%

Training Data Distribution

The model was trained on a balanced dataset with samples distributed across different scenarios and lighting conditions.

⚡ Performance Metrics

On-Device Performance

• Inference Time: 695ms
• Peak RAM Usage: 119.4K
• Flash Usage: 81.1K
• Overall Accuracy: 98.46%

Detailed Timing Breakdown

• Image Processing: 15ms latency, 4.0K RAM
• Object Detection: 1,085ms latency, 239.4K RAM
• Total Pipeline: 1,100ms

🔧 Hardware Requirements

Component	Details
ESP32-CAM Module	OV2640 camera sensor
OLED Display	128x64 resolution
MP1584 Buck Converter	Adjustable output for independent power supply
FTDI USB-to-Serial Converter	For development and programming
Prototyping PCB	For assembling components
Header Pins & Jumper Wires	For connections

Power Supply Mode	Description	Specifications
Development Mode	USB power via FTDI module	Flashing code via USB hub
Production Mode	MP1584 buck converter for standalone battery operation	Input: 4.5V–28V, Output: 5V adjustable, Up to 3A

📁 Project Structure

EdgeBlink/
├── img/                               # Documentation images
├── pcb/                               # PCB design files
│   ├── dct.kicad_prl                  # KiCad project files
│   ├── dct.kicad_pro                  # KiCad pro files
│   ├── dct.kicad_sch                  # Schematic files
│   └── fp-info-cache                  # Footprint cache
├── src/                               # Source code
│   ├── edge-impulse-sdk/              # Edge Impulse SDK
│   ├── model-parameters/              # Model configuration
│   ├── tflite-model/                  # TensorFlow Lite model
│   └── detect-esp32cam_inferencing.h  # Main inference header
│ 
├── detect.ino                         # Main Arduino sketch
├── library.properties                 # Library properties
└── README.md                          # This file

💻 Real-Time Detection Output

The system provides real-time detection feedback through serial output:

🔧 Model Optimization

The model uses several optimization techniques for ESP32 deployment:

EON™ Compiler Optimization

• 17% less RAM usage
• 36% less ROM usage
• Same accuracy maintained

Quantization Benefits

• Model Type: Quantized (int8)
• Size Reduction: Significant memory footprint reduction
• Speed Improvement: Faster inference times
• Accuracy Retention: No loss in detection accuracy

📈 Technical Specifications

Performance Characteristics

• Frame Rate: ~1.4 FPS (695ms inference + processing time)
• Resolution: Optimized for ESP32-CAM sensor capabilities
• Detection Range: Optimized for close-range object identification
• Power Consumption: Low power design suitable for battery operation

🙏 Acknowledgments

• Edge Impulse for the machine learning platform
• ESP32 community for hardware support
• TensorFlow Lite team for model optimization tools