Introduction. Traffic control devices, primarily traffic signs and road markings make the roadway environment intelligible to drivers and have a major impact on safety. Manual inspections remain the benchmark for compliance, but they are labour‑intensive, costly, and episodic, which creates long lag times between deterioration and remediation. Recent advances in computer vision and deep learning enable automated pipelines that detect, classify and assess the condition of signs using video and images, optionally supported by photometric measurements of retroreflectivity.
Problem statement. Despite high accuracy on public benchmarks, deep models degrade in real‑world edge cases: fading, dirt, graffiti, occlusions by foliage, snow or fog, and strong off‑axis viewpoints. Moreover, condition assessment requires quantified metrics — colour difference, glyph legibility and contrast, geometric deformation and retroreflectivity — rather than a coarse «good / damaged» label.
Purpose. To consolidate approaches for automated inspection, compare their strengths and limitations under realistic constraints, and outline an edge–cloud architecture that minimizes manual effort while meeting regulatory tolerances.
Materials and methods. We consider a spectrum of methods — from color/shape rules and classical hand-crafted features (HOG-SVM, SIFT-SVM) to single-stage object detectors (YOLOv8, SSD), two-stage detectors (Faster R-CNN, Mask R-CNN), and multi-task as well as multimodal approaches that incorporate depth maps (LiDAR/stereo). For quantitative condition assessment, we describe conversion to the CIELAB color space after Gray-World or Shades-of-Grey white balancing, the use of physics-guided networks to estimate retroreflectivity, and the evaluation metrics employed (mAP for recognition, RMSE for condition regression, FPS on Jetson-NX).
Results. Based on aggregated data, single- and two-stage deep detectors deliver mAP of 0.95 – 0.97, while multi-task/multimodal pipelines achieve the lowest error in condition estimation (RMSE 0.05 – 0.08). On edge devices, 18 – 35 fps is attainable (architecture-dependent), enabling on-device processing with subsequent offloading of candidate frames for heavy segmentation in the cloud. The proposed architecture combines a lightweight on-board YOLOv8-Nano detector (~28 fps) with cloud modules for segmentation and photometric analysis; contrastive pretraining on 20,000 unlabeled patches reduces labeling needs by ~60%, and an inexpensive solid-state LiDAR improves damage-class accuracy and enables tilt/roll measurement with ±2° precision.
Conclusions. AI‑assisted inspection substantially increases the frequency and objectivity of assessments, shortens maintenance cycles and creates the basis for data‑driven asset management. Future priorities include expanding open datasets with authentic degradation patterns, improving physics‑guided networks for direct retroreflectivity estimation, and conducting longitudinal studies to quantify safety and economic benefits.
