Literature Review on Road Damage Detection and Severity Recognition: Leveraging Computer Vision

Abstract

The increasing demand for quick and precise road maintenance has therefore been underscored by extensive research that was created over the last years, to find automated systems in which computer vision can work — harnessing deep learning methodologies to detect defects on roads & figure out how severe these actually are. It provides a discussion of different methodologies and technologies in these pipelines such as Convolutional Neural Networks (CNNs), YOLO models, ensembling learning etc. They have a good performance in various kinds of road damage: cracks, potholes, and surface deformations identification with high accuracy. They also use methods like image tiling, transfer learning, and multiple spectral data to increase their robustness so that models can be used effectively in different environmental conditions.But the avoidance of problems is still a long way off. Nonetheless, the disparate road scenarios across geographies and absence of labeled high-quality datasets remained as some common bottlenecks in addition to real-time-driven requirements that mandate significant compute powers. Although much more headway has been made in recognizing severity (using segmentation approaches and multimodal data, in particular), the combination of detection and classification altogether within a single joint pipeline is significantly harder.A comprehensive analysis of state‑of-the‑art software tools, recent challenges, and advances to improve the detection/ severity identification capabilities of road damage systems. It emphasizes how to address data problems, enhance model generalization capability in unseen scenes(a.k.a novel), and minimize latency. One final takeaway is the valuable reminder of the need to create operational standards for classifier performance required to drive practice adoption and scale in larger deployment settings ideally incentivizing public-private partnerships.