Optimizing YOLOv4 Hyperparameters for Enhanced Vehicle Detection in Intelligent Transportation Systems

Abstract Optimizing the hyperparameters of object detection models is critical for enhancing their performance, particularly towards domain-specific tasks such as vehicle detection. This work systematically investigates the optimization of key hyperparameters for the YOLOv4 model to maximize its efficiency in detecting vehicles. We evaluated the impact of two base CNN architectures (Tiny and CSP-DarkNet53), solvers (SGDM, RMSProp, Adam), learning rates (10 − 5 , 10 − 4 , 10 − 3 ), and detection thresholds (0.1, 0.2, 0.3) on the model performance. Our findings reveal that the optimal performance, achieving an average precision-recall (PR (avg)) value of 99%, is obtained using CSP-DarkNet53 network with a learning rate of 10 − 5 and a detection threshold of 0.1, employing the Adam solver across all learning rates and detection thresholds studied herein. Decreasing the learning rate from 10 − 3 to 10 − 5 and detection thresholds from 0.3 to 0.1, steadily enhanced network performance. Additionally, the choice of solver significantly influences the model’s performance. The results emphasize the importance of meticulous hyperparameter tuning to improve the accuracy of object detection models. By leveraging the optimized YOLOv4 network, traffic management authorities and autonomous vehicle developers can benefit from a more reliable and efficient vehicle monitoring solution. This study contributes to the broader field of intelligent transportation systems, offering a pathway to practical implementations of object detection technologies in real-world applications.