Detecting Road Defects and Hazards in Metropolitan Environments Using Optimized Deep Learning Techniques

Road networks are significant and instrumental assets for efficiently and reliably connecting last-mile distribution centers to end consumers. They play a crucial role in facilitating the final leg of the supply chain, ensuring timely deliveries and customer satisfaction. Road cracks and potholes on road networks pose significant challenges because they impede the smooth flow of goods and vehicles. These infrastructure issues can lead to delays, vehicle damage, and safety concerns, thereby increasing transportation and human costs. Delivery services between distribution centers and end users have recently utilized autonomous trucks and other delivery vehicles. Autonomous vehicle navigation systems need to be equipped with artificial intelligence models to timely detect and maneuver to avoid potential damage to the vehicle as well as the goods they are transporting. The deployment of autonomous delivery vehicles in metropolitan cities characterized by damaged roads poses a unique challenge. To ensure the safe and reliable provision of logistics services, autonomous vehicles must detect and avoid potholes and road cracks. In this paper, our research focuses on the critical task of object detection, with a specific emphasis on identifying potholes and road cracks, so that companies providing Mobility as a Service (MaaS) or autonomous vehicle companies may integrate object detection models into their systems. Leveraging advanced deep learning techniques, we propose a novel approach that utilizes an optimized object detection algorithm with spatial analysis to accurately detect and classify potholes and road cracks in real-time. In this study, we utilized the latest deep learning You Look Only Once (YOLO) models and optimized them for better results using hyperparameter techniques. To compare the performance of these optimized models, the mean average precision (mAP) at 50% intersection over union (mAP50) metric was selected. The results indicated the efficacy of the trained models in accurately identifying road irregularities amidst the complex road surface structure. Our findings have significant implications for the logistics industry, allowing carriers to employ autonomous vehicles to optimize delivery operations by detecting road irregularities and taking appropriate decisions to avoid them.