Empowering Road Safety: A Deep Learning Approach : Enhanced Unsupervised Road Anomaly Detection Using Deep Learning Models Leveraging Customized Dataset
Faroque, S M Adnan (2024)
Faroque, S M Adnan
2024
Tietojenkäsittelyopin maisteriohjelma - Master's Programme in Computer Science
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-12-30
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024122711691
https://urn.fi/URN:NBN:fi:tuni-2024122711691
Tiivistelmä
Anomaly detection applications have grown in various fields. Particularly its applications are observed in industrial fault detection. This study explores the application of anomaly detection for road safety. This thesis studies the performance of many anomaly detection models on a customgenerated road dataset with large anomalies and heterogeneous data. The working hypothesis is based on the assumption that SOTA models that work well with data that have high structural similarities and small anomaly regions will underperform in the case of roads that have highly varied structures and large anomalies. Comparative structural differences were analyzed between MVTec AD, BTAD, and the road dataset using SSIM. The performance of several SOTA models on the road dataset was assessed using pixel and image AUROC metrics. Csflow and Reverse Distillation methods performed better than the rest of the models with pixel AUROC 70% and 0.77% respectively on aggregated ’all’ category of the road dataset. The results show grounds for further work and room for improvement.