Optimization of SSD MobileNet Model for Road Damage Detection System Using Grid Search Hyperparameter Tuning

Road infrastructure plays a crucial role in land transportation, necessitating regular maintenance to prevent accidents, reduce vehicle damage, and ensure mobility efficiency. Traditional road condition assessments rely on high-end sensors and manual inspections, which are costly, time-consuming, and subjective. As a more efficient alternative, this study proposes an automated road damage detection system using the SSD MobileNetV2 model, optimized through hyperparameter tuning via the Grid Search method. The system leverages Deep Learning techniques and image-based inputs to classify three types of road surface damage: potholes, longitudinal cracks, and alligator cracks. Hyperparameter tuning focuses on selecting the most effective optimizer among Adam, SGD with Momentum, and RMSprop, with performance evaluated using accuracy, precision, recall, and F1-score. The final model is further fine-tuned and deployed on the Jetson AGX Orin embedded system for real-time road damage detection. Experiment in this research shows that the model, optimized using grid search for hyperparameter tuning, achieves an accuracy of 84.06%, a mean Average Precision (mAP) of 0.683 at an IoU threshold of 0.5, and a total loss of 0.1079. When deployed on the Jetson AGX Orin using video samples, the model achieves detection accuracies of 62.3% at ${1 0}\ \text{km} / \mathrm{h}, {5 6. 8 \%}$ at ${20\ \text{km}} / \mathrm{h}$, and ${5 2. 2 \%}$ at ${3 0\ \text{km}} / \mathrm{h}$.