Robust Optimization of SSD for Real-Time Pothole Detection Under Variable Road Conditions

Abstract

Road conditions significantly affect the safety and efficiency of transportation, and problematic roads are a challenge because potholes are unpredictable. There are a number of inherent problems with existing pothole detection techniques, namely manual inspection, which share common vices of being time-consuming, expensive, and unable to be adopted on a large scale. This work introduces the real-time pothole detection and maintenance integration of a Smart Road Monitoring System using deep learning SSD. A comparison of the SSD300 model and the proposed Mask R-CNN and YOLOv8 algorithms is done on measurement indicators such as mAP, precision, recall, F1-score, and inference time. The experimental outcomes indicate that, for all scales and aspect ratios and with a higher inference time within the desired range, SSD300 yields the highest mAP of 78.2%, therefore being the most suitable for real-time implementations. YOLOv8 has a comparatively low accuracy of 85.3% mAP, yet it is more accurate and faster than its counterparts, with only 18.2 ms to complete, although it all comes with a higher precedent, thus not as scalable. While Mask R-CNN achieved a high precision of 82.4% it has a relatively slow inference time of 130.6 milliseconds. When installed on an SSD, the new pothole detection system excels in being able to detect potholes and provide reports and solutions without needing many human interventions, thus increasing road safety and maintenance. This research shows that SSD is feasible for monitoring smart roads and is relatively efficient for constructing a scalable solution. Future work will extend the presented results with additional data, IoT, and more complex predictive maintenance in order to optimize the management of infrastructures and traffic networks.