Deep Learning Techniques for Street Sign Identification in Urban Environments

Identifying small objects, such as street signs in street environments, is a challenging task due to varying lighting conditions, occlusions, and environmental factors. Several approaches have already been proposed regarding the same. Although these approaches provides descent performance yet fail to provide optimal performance in the low lighting conditions. This paper proposes a novel graphical approach designed to provide more accurate classification in normal and low lighting conditions also. In the proposed algorithm, features are extracted using simplified Gabor feature maps. It provides more optimal and realistic features to the CNNs for more optimal classifications. By using consistent parameters across both detection and classification stages, the system achieves smooth and precise performance. Proposed algorithm is trained on a comprehensive dataset comprising over 50,000 images of street signs across 43 distinct classes. This extensive training enables the system to handle various real-world challenges, including changes in distance, orientation, and partial visibility. The approach has demonstrated significant robustness in diverse scenarios, achieving an impressive accuracy of 97% under favorable daylight conditions. However, the model's performance declines to 74% in extreme lowlight night conditions, revealing the need for further optimization to address low-light challenges. This work provides a strong foundation for efficient and accurate small-object detection, highlighting areas for future improvement. By addressing lowlight performance, the system could further enhance its applications in autonomous driving, traffic management, and urban navigation systems.