International Journal of Science, Engineering and TechnologyAuthors: Tambe Harshal Maruti, Professor A.P. Banger, Professor S. B. Bhosale, Professor V. A. Karad
Abstract: Potholes wear out vehicles, hold up traffic and cause crashes, yet manual road inspection is slow, expensive and cannot provide a continuous, geotagged record. This paper is a focused narrative review of five IEEE and IET studies on automated pothole detection: a comparison of YOLOv3, SSD, HOG with SVM and Faster R-CNN; a study that fine-tunes four deep detectors on road images from around the world and runs the chosen detector on a real vehicle that sends alerts to an Internet of Things platform; a convolutional-network prototype for autonomous vehicles; an assessment of R-CNN and SSD MobileNet through an object detection API; and a ResNet-50 system that detects potholes and estimates their depth. The review looks at three linked themes: the sensing approaches and detection models in use, how strong and how comparable the reported evidence is, and what matters when detection is stretched into a reporting pipeline. Only one of the studies could be read in full. It reports a mean average precision of 82.02% (IoU 0.4) for Faster R-CNN ResNet101 and 5–6 frames per second for a lighter model on an embedded automotive platform. The other four were judged from their published abstracts alone; among other figures, they give 99.02% accuracy for a prototype CNN and a maximum detection confidence of 93%. Because these numbers rest on different data and metrics, they cannot be ranked against one another. Just one study carries detections onward to a cloud platform. The review offers a comparative table, a conceptual architecture for the proposed pothole detection and reporting system (not experimentally validated) and a prioritised list of research gaps.
