In New Zealand alone more than 3000 vehicles are currently reported as being stolen.
Identifying data for stolen cars is currently made public through through the police.
Handheld devices nowadays feature the computational capababilities, as well as optical
sensors, geo-location and internet connectivity.
Our idea is to crowd-source the localization of stolen vehicles with smartphone devices.
This way the police and insurance company will be able to search for stolen vehicles with the help of citizens.
Our working prototype takes in video footage and uses a state-of-the art real-time deep
learning algorithm to detect license plates and runs character recognition on it.
We query that data against the current police dataset and inform the police about a
positive match.
Our application queries the police database, downloads die information in the given csv
format and stores the data on stolen vehicles on the users device.
While driving, the camera is used to run a two-stage algorithm to read license plates.
The first stage utilizes the tiny yolo / darknet framework, which is a deep learning
convolutional neural network, that we custom trained with a self-annotated dataset,
based on dash-cam footage. This stage creates bounding boxes for possible number plate
positions.
The second stage utilizes OpenAPNR to run OCR on the detection image sections and reads
the text on the number plate. The resulting text is queried against the locally stored
database of stolen vehicles and sends the license plate and current GPS coordinates
to the backend, which can visualize the geo-location data via the HERE api.
Our software was build in Python and C++ with the help of libraries like OpenCV, darknet,
OpenANPR and cuda.
For demonstration purpose we also create a small web server that can be used for query stolen car information. The server can find all locations where particular stolen car was seen recently and police officer or insurance company stuff can click on every marker and make sure that it was actually the car that was stolen. Such visualization can also help to identify certain patterns in thief behaviour and help to arrest the thief.
