The Light-Weighted Road Fault Identification for Embedded Hardware Platform Using YOLO

Maintaining roads in Indonesia is challenging due to increasing road defects and rising accidents. Manual maintenance is time-consuming, labor-intensive, and prone to human error. Deep Learning, particularly Convolutional Neural Networks (CNNs), shows promise but lacks sufficient training datasets. Based on these shortcomings and limitations, this research proposes several contributions to address the existing issues. First, we introduce a new dataset comprising thousands of images featuring three common road faults in Indonesia: potholes, cracks, and alligators. The dataset is annotated and utilized to train both default YOLO model and fine-tuned model in terms of road-faults detection tasks. Second, this study presents a fine-tuned model of YOLO using sparsity training, achieving higher detection accuracy while minimizing computational costs, reducing detection time, and ensuring efficient and reliable detection of road faults. Third, the study embeds the model into a small hardware platform by converting the weights into different frameworks, demonstrating its lightweight nature, low power consumption, compact size, and high reliability. Evaluation metrics, including mean Average Precision (mAP), precision, and recall are employed to assess model performance, with satisfactory results when tested on various road images, effectively identifying faults.