Computer Vision-Based Detection and Classification of Road Obstacles: Systematic Literature Review

Road obstacles are a major contributor to traffic accidents, making their accurate detection and classification vital for road safety and infrastructure maintenance. This paper presents a systematic literature review (SLR) of 95 peer-reviewed studies published between 2015 and 2024, specifically focusing on computer vision-based approaches to detecting and classifying road anomalies such as potholes, cracks, manhole covers, and speed bumps. Potholes were the most frequently studied obstacle (56 studies), followed by cracks (19), while manholes and speed bumps were the least represented (6 each). Detection tasks dominated (44.29%), with classification (30%) and combined approaches (25.75%) also explored. Deep learning techniques, particularly YOLO variants and CNNs, were the most used (87.14%), achieving state-of-the-art performance—e.g., YOLOv8 attained 100% accuracy and 99.50% mAP on benchmark datasets. Public datasets were used in 36.59% of cases, and 91.55% of studies relied on image-based inputs. Preprocessing techniques like data augmentation (30.86%) and annotation (24.69%) were prevalent. Although this review strictly focused on computer vision techniques, it identifies future research directions such as multi-sensor fusion, machine learning forecasting, crowdsourced data integration, and smart city infrastructure alignment. These broader approaches, though beyond this review’s scope, are critical for developing more comprehensive, real-time, and context-aware road obstacle detection systems.