Developing ML-based Automated License Plate Recognition (ALPR) software for an American surveillance technology provider.

The Daffodil team set about developing the ML-based web and mobile application that could eventually scan the license plates of fast-moving vehicles and retrieve valuable information about the vehicles and its owner. This was supposed to provide insights to law enforcement personnel to enable them to solve crimes more accurately using their existing mobile devices and stationary cameras. The following capabilities were then worked into the mobile and web-based application built on an ML foundation:

Custom ML Model

The team leveraged the TensorFlow platform to create and train the ML model with custom training loops. In the ALPR pipeline, automobiles are first located in the frame using an object detection deep learning model, the license plate is then localized using a license plate detection model, and finally, the characters on the license plate are recognized. Eighty percent of the data is used for the training set and twenty percent is used for the validation set. The dataset is then randomly split into the necessary number of records using a series of instructions.

Instantaneous Data Retrieval

When a wanted plate is seen, the automatic system captures, compares, and analyses vehicle license plates and notifies the closest law enforcement agency or sheriff’s department within seconds. The system checks to make sure the information is accurate, the license plate is from the same state as the one where the vehicle is wanted, the reason this car or its owner was wanted or of interest, and if the case is still relevant. The searches can be made on the model, make and color with instantaneous data retrieval.

Cost-Effective ML Validation

Client-side validation was carried out in TensorFlow wherein the application developed by the team used ML directly on the client side. Instead of waiting for API results about the vehicle, client-side computing achieves analysis of live feed and the ML model inference in real time. There is also a reduction in costs since you no longer need to maintain ML servers for heavy computations.

Incrementally Improving Performance

The YOLO v2 real-time object detection convolutional model has been the basis for the creation of the custom law enforcement ALPR model created by team Daffodil. It enables the model to detect small objects in large groups, accounting for localization errors, with a high-resolution classifier. With a single forward propagation through a neural network, the prediction in the entire image is done in a single algorithm run, which helps reduce the computation time.