iYOLOV7-TPE-SS: Leveraging Improved YOLO Model With Multilevel Hyperparameter Optimization for Road Damage Detection on Edge Devices

Digital twin cities integrated road damage inspection to support smart road infrastructure. The road’s condition was assessed using road damage survey metrics, and the issue was resolved manually. Real-time, automatically, and precisely employing artificial intelligence (AI) is needed to improve road safety and reduce maintenance costs. This research on road defect detection utilizes improved YOLOV7 (iYOLOV7) while multilevel hyperparameter optimization using combination of the Tree-Structured Parzen Estimator (TPE) and Search Space (SS). This model we call TPE-SS. The iYOLOV7 provides real-time processing on edge devices that have limited storage, low-speed processors, and memory. The research investigates how leveraging multilevel hyperparameter optimization using the TPE-SS model improves system performance where road damage is detected. The results of iYOLOV7 model demonstrate the smallest total loss of 0.1, indicating a significant performance improvement where a precision value of 0.986, an average recall (AR) of 0.970, a mean average precision [email protected] of 0.988, a mean average precision [email protected] of 0.806, and an F1-score of 0.978. The measurement results and analysis findings reveal that our proposed method is accurate enough. The embedded system employs NVIDIA Jetson Nano for inference, which takes only 0.135 seconds and has a scalability performance of 7.470 FPS to recognize things associated with road damage. Although the employed NVIDIA AGX Orin edge device may produce a higher FPS of 67.034 and a faster inference time of 0.014 seconds. This means that the iYOLOV7 model is extremely feasible and practical for usage on edge devices to identify road damage.