Analýza a validace vizuálního stavu dopravního značení

Abstract

This bachelor thesis deals with the design and implementation of a two-stage system for vertical traffic sign detection and visual condition assessment. The first stage employs a YOLO11l detector fine-tuned on the GTSDB dataset augmented with synthetically generated Czech traffic sign data, covering a total of 57 classes. In the second stage, the detected sign crop is evaluated in terms of visual condition across three categories: good, faded, and degraded. Three approaches to condition assessment were implemented and compared: a reference-based approach using comparison with a template image, a classification approach using an EfficientNet-B0 network, and a hybrid approach combining both. The detector achieved a mAP50 of 0.978 on the validation set, while the condition classifier reached an overall accuracy of 98.2 %. The system was verified on real dashcam recordings and processes approximately 20 frames per second.