Vehicle-Mounted Vision Sensing for Large-Scale Road Surface Monitoring: A Multi-Year Field Study of Acquisition...

Vehicle-mounted vision sensing, in which cameras on patrol vehicles acquire ground-level imagery across a road network, offers a low-cost, scalable approach to large-scale road surface monitoring and pothole detection. However, its real-world effectiveness depends not only on detector accuracy on curated imagery but also on two rarely reported factors: field acquisition conditions and pipeline architecture. We present a multi-year field study of a vehicle-based pothole-monitoring system. First, controlled field comparisons show how acquisition geometry and device thermal behavior affect detection yield: across three vehicle types, an actively cooled windshield mount achieved yields of 41.7-58.3%, versus 0-4.2% for a rear-view-mirror baseline. Second, we identify a shared-bias limitation in a homogeneous two-stage detector cascade, in which a second You Only Look Once (YOLO) detector reproduces rather than eliminates first-stage false positives. Third, we redesign the pipeline as a heterogeneous detection-to-classification cascade: an edge-deployed YOLOv12-large detector proposes candidate regions, and a server-side EfficientNet-B2 classifier verifies each. On field data, this raised the detection mean average precision (mAP) from 32.7% to 69.5% and reduced the false-positive rate from 99.96% to 32.26%, cutting manual inspection workload by about 70%. Unlike prior work that mainly optimizes detector architectures, this study shows that acquisition geometry and heterogeneous verification dominate real-world sensing performance under operational deployment.