NaijaTrafficNet: A Custom CNN for Robust Sign Classification under Nigerian Road Conditions

In this study, we propose NaijaTrafficNet, a novel custom Convolutional Neural Network (CNN) architecture designed for robust traffic sign classification under challenging Nigerian road conditions, characterised by poor lighting, occlusion, and non-standard signage. To address data scarcity, we merged two public datasets the Kaggle Traffic Sign Dataset Classification and the German Traffic Sign Recognition Benchmark (GTSRB) resulting in a unified dataset of 12,472 images across 43 classes. Extensive preprocessing, including resizing, grayscale conversion, histogram equalization, and data augmentation, was applied to enhance generalizability. The model was trained from scratch using the Adam optimizer and evaluated on a held-out test set of 795 images. NaijaTrafficNet achieved a test accuracy of 94.72% and a weighted F1-score of 94.48%, demonstrating high performance, particularly for regulatory signs (e.g., stop and speed limits). The architecture is lightweight, enabling real-time inference suitable for deployment in resource-constrained environments. Limitations include misclassification of visually similar signs (e.g., 50 km/h vs. 60 km/h). This work contributes (1) an open-source preprocessing pipeline for African traffic sign data, (2) a novel CNN architecture, and (3) empirical validation of deep learning for non-standard signage. Future work includes comparative benchmarking and synthetic data generation.