License Plate Detection and Recognition in Video Surveillance Sequences

Over the past few years, urban security and optimized traffic management have driven the demand for increasingly intelligent and efficient surveillance systems. While Automatic License Plate Recognition (ALPR) is central to these objectives, recognizing license plates in real-world surveillance videos remains a significant challenge. Environmental factors such as motion blur, poor illumination, and complex camera angles severely complicate accurate detection and decoding. The integration of robust, automated approaches into video analysis is thus essential to maximize system reliability. This master's thesis proposes a comprehensive framework for automated license plate detection and recognition applied to video surveillance streams. To achieve this, we develop a system that integrates the RF-DETR neural network with classical computer vision techniques. Specifically, the deep learning architecture is utilized for the accurate spatial segmentation of license plates within video frames. This is then complemented by a dual Convolutional Neural Network (CNN) system for character extraction: a first denoising network is employed to restore degraded plate images, while a second prediction network extracts the characters by leveraging both the original and the restored visual data. Finally, we evaluate our framework against a custom-annotated ground truth dataset derived from real-world security footage. System performance is assessed in terms of detection precision, character recognition accuracy, and real-time processing capabilities, demonstrating that the synergy between deep learning models and classical techniques provides a highly robust solution for complex operational contexts.