This repository contains the entire code used to developed a car plate detection system, from training and testing the model, to writing scripts to run in Raspberry Pi Model 4-B. First selecting an appropriate dataset from Kaggle, which includes images along with essential metadata like image dimensions and bounding box coordinates in .xml format. For model selection, after trying and testing various CNN architectures, Faster-RCNN combined with MobileNet V2 offered the best balance between performance and memory requirements. Faster-RCNN has a significant performance boost due to the integrated Region Proposal Network, while MobileNet V2 contributed higher resolution feature maps, enhancing accuracy. Hardware integration involved connecting a camera module via a MIPI CSI-2 port, downloading necessary libraries using pip, enabling the camera through the configuration menu, and feeding captured images into the model for inference.
- Insert the Micro SD Card with the Raspberry Pi OS image.
- Connect the Camera module to the MIPI CSI - 2 connector.
- Connect the required external output devices like monitor, keyboard and mouse to the Raspberry Pi.
- Directly connect the USB-C cable for power supply after ensuring there are no loose connections.
- After the desktop loads, run the code for model inference, image capture and backend communication.
Note: Burn a Raspberry Pi OS Image using the instructions given in this link.
Install the below libraries for efficient working of code:
pip install pynput picamera2 numpy matplotlib torch torchvision pytesseract easyocr firebase-admin opencv-python
For the GPU version of pytorch, run:
pip uninstall torch torch audio torchvision
pip cache purge
pip install torch torchaudio torchvision -f https://download.pytorch.org/whl/torch_stable.html
