Work

License plate detection, on vehicles traveling at up to 100 mph, based on morphological operators.

Low resolution images:

High resolution images:

Sample detected license plate:

Sample segmented characters:

Character segmentation

Shadow suppression

License plate skew and orientation normalization using the Radon Transform

SOM on character features

Make and Model Recognition (MMR) using the Scale Invariant Feature Transform (SIFT)

The algorithm matches the set of detected feature points belonging to one car with the detected feature points of another. The more points that match between the two cars, the higher the probability is that they are of the same make and model.

A database of vehicles of known make and models is to be constructed. These are to be called templates or template vehicles. Every entry of the database is to contain one of these vehicles. It should be tested if it is beneficial to store multiple vehicles, or templates, of the same make and models to account for possible alterations and/or modifications that the query vehicle may contain.

The algorithm should be applied to match a query vehicle, whose make-and-model is to be identified, with a template vehicle, whose make and model is known. The template’s make and model which matches, with the highest probability, the query vehicle will determine the make and model of the query vehicle.

This image shows two vehicles whose make and model match.

This image shows two vehicles whose make and models do not match.

This image shows two vehicles whose make and models match.

Cargo Container Identifier ROI Finder based on morphological operators

Back door bars (obstructions) removal

Truck license plate ROI Detection

Cargo Container Orientation Classification

Original Images – Container rear view

Pre-processed Images – Binary images, preprocessed with morphological operators

Original Images – Container front view

Pre-processed Images – Binary images, preprocessed with morphological operators

The preprocessed images, similar to the above, were then coded as binary feature vectors and used for training. Feature extraction and construction of feature vectors:

A Support Vector Machine was trained with 200 pairs cargo container images (front and rear). Then, with another set of 200 pairs of cargo container images (front and rear) the classifier was tested and achieved 100% accuracy.

Further testing needs to be performed with, for example, container images under varied lighting conditions, fields of view, conditions of the containers (damaged and not damaged), etc.

EasyMatch: Image Difference for hidden object detection