Robust Road Sign Feature Extraction Through Data Curation and Multi-Task Learning for Global Map ‎Creation

Abstract

This research presents a comprehensive approach to enhancing road sign feature extraction for global map creation through ‎strategic improvements in data quality, feature learning, and network architecture. Designed to address core challenges in HERE ‎Technologies' map creation pipeline (US patent PAN: 18/988231), our approach significantly improves the Stage 1 component ‎of their patented three-stage framework by replacing the previous YOLOv7-based implementation with a more robust and ‎effective solution. The methodology centers on three key innovations: (1) an intelligent data curation and filtering strategy that ‎reduces annotation noise by 37% and improves overall data quality without extensive manual re-annotation; (2) novel self-‎supervised pretext tasks that develop rich feature representations of road sign characteristics such as color, shape, and ‎contextual positioning; and (3) a multi-headed network architecture that preserves geometric understanding while enabling ‎simultaneous optimization of detection, segmentation, and classification tasks. These innovations collectively address critical ‎map creation challenges, including domain divergence between different imagery sources, class imbalance across sign types, ‎data scarcity for rare classes, and noisy training samples. Evaluation metrics demonstrate exceptional improvements, with the ‎enhanced system achieving 92% precision, 93% mAP@0.5 for detection, and processing inputs 64.29% faster than the ‎previous implementation while simultaneously performing multiple tasks. The approach significantly improves performance in ‎challenging scenarios, with a 53% improvement in adverse lighting conditions and 31% higher accuracy in poor weather. By ‎focusing on fundamental improvements in data quality, feature representation, and architectural design rather than simply ‎adopting newer base models, this work establishes a foundation for more efficient and accurate feature extraction that enables ‎faster global expansion of map coverage without sacrificing quality‎.