Improving Road Safety

UAE law enforcement needed to tackle traffic violations in an automated way. The task was to analyze video streams, identify certain violations, read the offender’s license plate and alert police, without using any server-side APIs. Existing off-the-shelf solutions were inaccurate for the specific needs and the Emirates region, and also weren’t optimized for edge computing.

We developed a fully customized pipeline consisting of ALPR (automatic license plate recognition) and specific violation detection models, and optimized the pipeline for Nvidia TX2 devices in order to be independent on internet connectivity. We also created a robust training data generator to increase the accuracy of neural networks using a combination of real-world and synthetic data. The resulting accuracy was substantially higher than that of the commercially available solutions.