Vision-based Smart Parking System

Hardware Setup
We kicked off the project by building a simple and stable hardware system using the ESP32-CAM module paired with a 5MP OV5640 camera. For power, we used dual IMR18650 batteries (7.4V), stabilized with an LM7805 voltage regulator and a 3.3μF capacitor to reduce noise. The output provides a clean 5V to the ESP32-CAM’s VCC and GND pins. This minimal setup ensures the ESP32 can consistently capture images without power fluctuations.

Dataset Collection & Labeling
Images were captured using the OV5640 camera at 160×120 resolution in various lighting conditions (morning, afternoon, evening). A total of 181 images were collected. Each image was annotated using LabelImg with Pascal VOC format. Parking slots were labeled with bounding boxes as either empty (class 0) or occupied (class 1) — resulting in 2,896 labeled objects. This diverse dataset was key for a robust model.

AI Model Training
The dataset was trained using Edge Impulse with FOMO MobileNetV2 0.35, a lightweight architecture ideal for edge deployment. Images were resized to 160×160 and normalized. Training ran for 60 epochs with a learning rate of 0.001 and batch size of 32. The data split was 81% for training and 19% for testing. The result? 97.14% accuracy and 98.7% F1-score on test data — more than good enough for real-time parking detection.

Model Deployment to ESP32-CAM
After training, the model was exported in TensorFlow Lite (.tflite) format, then compiled to a C++ header (.h) using Edge Impulse’s EON Compiler. We flashed it to the ESP32-CAM using Arduino IDE and the esp32_camera example as base code. The model footprint is just 626.5KB and can infer a frame in around 2.3 seconds, directly on the ESP32 — no cloud needed!

Cloud Infrastructure Setup
We built a secure and scalable cloud backend using AWS. The Laravel backend runs on EC2 (t2.micro, Ubuntu 22.04) in a public subnet with Elastic IP. MySQL 8.0 runs on Amazon RDS in a private subnet, protected by security groups. Prediction tasks are automated using AWS Lambda (Python), triggered by EventBridge every 30 days. It’s fully serverless and cost-efficient.

Backend Development
Our backend uses Laravel 10 (PHP 8.2), featuring a dedicated endpoint /live-parking-slot to receive ESP32-CAM data via POST. Data includes the timestamp, parking slot status, and CRC32 hash for integrity. It’s stored in the live_parking_prediction table. The API is designed with RESTful principles for seamless integration with any dashboard or frontend client.

Data Transmission
Every 10 seconds, the ESP32-CAM sends a JSON payload to the server. This includes the real-time status of 16 parking slots in binary format (e.g. 001011001...), a timestamp, and a checksum. The data is pushed to http://54.86.233.199/live-parking-slot using the built-in HTTPClient library. It’s fast, light, and reliable — even on limited bandwidth.

Automated Parking Forecasting
Using AWS Lambda, we fetch 30 days of historical parking data and train a Prophet time series model. The forecast includes hourly predictions from 6 AM to 9 PM and is saved as a CSV string in the parking_prediction table. These predictions help users plan ahead and avoid full parking lots — powered entirely by serverless AI.

Frontend Interface
The web interface is built with Laravel Blade and Tailwind CSS for a clean, responsive mobile UI. The real-time parking status is displayed in a 4×4 slot grid — with white for empty, red for occupied — updated every few seconds via AJAX. Users can also view hourly prediction charts using Chart.js to see expected parking load throughout the day.

Testing & Optimization
We validated the system using over 50 test scenarios with toy cars under different lighting. The goal was at least 95% accuracy, and we exceeded that. End-to-end response time (from image capture to dashboard update) was under 8 seconds, measured with Wireshark. A 72-hour stress test using Locust (1,000 req/min) achieved 99.98% uptime, and memory optimization reduced RAM usage by 12% on the ESP32.