Abstract:
The durability and safety of road infrastructure are essential for economic activity and
public safety. However, the traditional manual inspection of road surfaces for damage
is often inefficient, costly, and susceptible to human error, making automated solutions
highly desirable. This research explores the potential of unsupervised anomaly detection
techniques, namely Variational Autoencoders (VAEs) and Isolation Forests (iForest), for
automating road damage detection and severity estimation. By training these models
solely on undamaged road images, this study leverages latent space representations from
VAEs and anomaly scoring from iForest to detect deviations indicative of road damage.
Testing was conducted on a diverse dataset of road images, allowing the models to identify
damage and classify its severity into three categories: low, medium, and high. The study
also compares these results with those of the RT-DSAFDet (Dynamic Scale-Aware Fusion
Detection) model. While iForest and VAE achieved high recall rates and performed well
in severity classification, RT-DSAFDet demonstrated superior adaptability to complex
road scenarios and lighting conditions, making it well-suited for real-time applications.
ii
