Deep Learning Models for Traffic Sign Recognition in Ecuador: Challenges and Solutions

This study focuses on developing and evaluating deep learning models for traffic sign recognition in Ecuador. Given the unique characteristics of Ecuadorian traffic signs, particularly the “yield” sign with the phrase “Ceda el Paso”, this research addresses the challenges of adapting existing datasets and creating new data specific to the region. The study uses convolutional neural networks (CNNs) to improve the recognition and classification of four primary types of traffic signs: “stop”, “yield”, traffic lights, and speed limits. The methodology involved modifying the CRISP-DM approach to better suit academic research by focusing on relevant objectives and excluding the deployment phase due to the model’s specific scope. Data were sourced from public datasets and manually collected pictures, leading to a comprehensive database of 5322 pictures for training and validation and 400 pictures for testing. Data augmentation and preprocessing techniques were employed to enhance model performance. Three models were developed and tested: a traditional neural network, a CNN without regularization, and a CNN with regularization. The regularized CNN model performed best, achieving 91% accuracy on validation data and 81% on test data, demonstrating its robustness and generalizability. The study highlights the importance of having a broad and representative dataset and the effectiveness of regularization techniques in preventing overfitting.