Design of a Background Extraction Platform with Semantic Elements

This work presents the design and implementation of a modular software platform for semantic segmentation and 3D reconstruction of road scenes using monocular vision and GPS data. The system focuses on extracting vertical traffic signs and background information from road images, generating georeferenced point clouds in. pcd format. The proposed method leverages a fine-tuned Dense Prediction Transformer (DPT) for semantic segmentation. The platform processes synchronized video and GPS data, applies inference on each frame, and fuses depth, segmantic data, and location information to build a spatially 3D representation of the environment. The fine-tuned model achieved high performance, with a maximum Intersection over Union (IoU) of 0.9447 and F1-score of 0.9709. The results demonstrate substantial improvements over the pre-trained baseline and help to give a field of expertise to the model.