High performance automatic number plate recognition in video streams

Abstract

We present a range of image and video analysis techniques that we have developed in connection with license plate recognition. Our methods focus on two areas - efficient image preprocessing to improve low-quality detection rate and combining the detection results from multiple frames to improve the accuracy of the recognized license plates. To evaluate our algorithms, we have implemented a complete ANPR system that detects and reads license plates. The system can process up to 110 frames per second on single CPU core and scales well to at least 4 cores. The recognition rate varies depending on the quality of video streams (amount of motion blur, resolution), but approaches 100% for clear, sharp license plate input data. The software is currently marketed commercially as CarID1. Some of our methods are more general and may have applications outside of the ANPR domain.