Detektering av förändring i trafikmärken med hjälp av artificiell intelligens och temporal data från mobil kartläggning

Efficient change detection of critical road assets like traffic signs enables structured planning, monitoring and maintenance (Hebel et al., 2014). Traffic signs frequently become damaged by accidents, vandalism or wear and tear. They may also become occluded by vegetation or other structures. Consequently, effective monitoring of changes in traffic signs enable authorities to improve traffic safety by reacting to changes in a timely manner (Creusen et al., 2012; Khataan and Gargoum, 2024). However, research on this topic remain limited, particularly with respect to automatic and deep learning methods that can reduce or replace time-consuming manual tasks such as traffic sign recognition or change labelling. This thesis develops and presents a pipeline for automatic change detection on traffic signs using temporal data collected by a mobile mapping system (MMS) mounted on a car. The proposed pipeline integrates existing and trained deep learning methods with new algorithms that perform change detection. It consists of traffic sign detection using the You Only Look Once (YOLO) model (Redmon et al., 2016), sign classification using the EfficientNetV2S architecture (Tan and Le, 2021), traffic sign positioning through triangulation and finally, change detection implemented with a Python algorithm. The resulting output is an inventory in which traffic sign changes are labelled along with their change type and additional relevant information, such as sign position and class. The results of the proposed change detection pipeline records change correctly for 84 % of the observations. Furthermore, this thesis highlights the value of proper training with well-balanced training datasets for the deep learning models used for traffic sign recognition. It also emphasised the importance of high-quality data for achieving reliable performance of the change detection pipeline.