License Plate detection using computer vision

💡 Inspiration 💡

There are many types of vehicle license plates in India, including new energy license plates, large truck license plates, government vehicle license plates, and military license plates. The existing commercial license plate recognition system only targets common license plates and does not completely cover the full range of license plates. Therefore, we come up with this model which will detect the bounding box in passenger cars and SUVs specific to India and if possible also recognize the text in the license plate.

❓ What it does ❓

With the advancement of industrialization, vehicles have become the preferred means of transportation for people to go out. There are also higher requirements for the task of license plate recognition. The license plate recognition is mainly divided into two parts, one is to accurately locate the license plate in the picture, and the other is to perform character recognition on the positioned license plate.

⚙️ How we built it ⚙️

From the first step, we collected the Indian vehicles with license plate with different sizes, different characters and possibly containing distortion, tilt, blur other noises. These are open source images which we picked up from GitHub. Then we labelled the images using Microsoft VoTT to export the annotations in csv. This was then loaded onto a computer vision program where we read the annotation and images together and then fed to VFNet model with swin transformer as backbone. The model has 70% coco metric accuracy. This model was then deployed on Habana AWS EC2 DL1 instance and loaded for inference

🚧 Challenges we ran into 🚧

Habana SDK is not simple to install and the instructions are not very clear. We essentially then created the model with pytorch framework and then deployed it on HABANA using their frameworks