Enhancing Vehicle Identification in Challenging Conditions Through Fine-Grained Classification

Automatic License Plate Recognition (ALPR)-based Vehicle Identification Systems are vital for modern traffic management and law enforcement, yet they frequently encounter challenges in real-world scenarios, leading to recognition errors. Our ongoing research investigates a novel method to improve ALPR accuracy through the integration of Fine-grained Vehicle Classification (FGVC), with a particular focus on vehicle make identification. By cross-referencing identified vehicle brands with registered data, we aim to enhance the reliability of ALPR-based vehicle identification systems. Nonetheless, initially, our work-in-progress is concentrated on refining FGVC techniques to facilitate their integration with ALPR. We assess four deep learning models for vehicle make classification and explore methods such as selective prediction and a new class reduction approach. Preliminary results from our experiments on a modified Brazilian vehicle dataset indicate that combining these methods significantly boosts vehicle make identification accuracy. This improved classification approach is anticipated to reduce false positives and increase recognition rates under challenging conditions. Future efforts are going to be directed towards integrating these observed enhancements into ALPR-based vehicle identification systems to further improve their performance in real-world applications.