COMPUTER VISION BASED AUTOMATED CRACK AND POTHOLE DETECTION IN PAVEMENTS USING DEEP LEARNING

Abstract

Road condition monitoring (RCM) or pavement distress detection has been a demanding strategic research area in maintaining a large network of transport infrastructure. With developments in vision-based and data handling techniques, in integration with high computational resources, numerous ground-breaking pavement distress assessment systems have been established in past few years. Majority of these technologies employ state-of-the-art sensors and vision-based artificial intelligence (AI) practices to assess, categorize and localize distresses in pavement using the measured data. The next-generation sensors (non-contact and contact) in addition to various methodologies and innovative contributions proposed by various researchers in this field promise a futuristic insight to transport infrastructure owners. The decisive role played by smart sensors and data acquisition platforms such as smartphones, drones, vehicles integrated with non-intrusive sensors such as RGB and thermal cameras, laser and GPR sensors in the performance of the system also needs special mention. Apart from sensing, development of AI technologies paves the way for a well-directed and all-inclusive futuristic research in pavement distress detection. Automated detection of pavement distress, driven by AI-assisted engineering solutions, promise an effective solution to prevent deterioration of premature surface disintegration in pavements. Pavements during their service life are subjected to varying conditions. With a booming world population exerting extra service load coupled with adverse climatic changes, ageing pavements, poor construction and design resulting in poor drainage systems are some of the major contributing factors to distress in pavements. These distresses appear as cracks, potholes, surface disintegration and deformations and are a severe threat to road safety and adversely affect the health of the pavement infrastructure. Thus, effective pavement maintenance strategies in the form of regular pavement monitoring are required. Regular pavement monitoring is a plausible solution only if the monitoring process is automated. To facilitate automation of the distress detection system, a computationally light and feasible, intelligent pavement distress detection system is proposed in this research study. In this research, a novel workflow is developed for image-based distress detection, which is further extended by integrating it with distress severity assessment. A single-stage CNN architecture, PC Retinanet is proposed which is obtained by modifying and optimizing the RetinaNet neural network through a series of trials and errors to best detect cracks and potholes.