Semantic Constellation Place Recognition Algorithm Based on Scene Text

• Published in: IEEE transactions on instrumentation and measurement Vol. 74; pp. 1 - 9
• Main Authors: Wen, Shuhuan, Long, Yidan, Li, Peng, Wang, Bohan, Qiu, Tony Z.
• Format: Journal Article
• Language: English
• Published: New York IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Character recognition; Computer vision; Feature extraction; Graph matching; Image recognition; Invariance; Multivehicle system; place recognition; position correction; Recognition; removal of occluded text; Robustness; semantic constellation matching; Semantics; Statistical analysis; Text detection; Text recognition; Three-dimensional displays; Training
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2025.3545984

Visual place recognition (VPR) represents a significant challenge within the domains of computer vision and autonomous vehicles. Due to the dynamic nature of real-world environments, ensuring that the accuracy and robustness of place recognition remains a critical concern. Scene text exhibits rotation invariance and maintains its uniqueness and invariance even under significant changes in environmental factors such as lighting and viewpoint. This article proposes a novel semantic constellation place recognition algorithm that leverages the unique properties of scene text. The algorithm initiates by detecting and recognizing scene text, followed by the removal of occluded text through statistical analysis of text occurrence frequencies and text similarity computation. To rectify the positional information of irregular scene text, the intersection points of text detection box diagonals are replaced with speeded-up robust features (SURF) points within the text detection box, thereby representing the complete text position. Furthermore, descriptors based on the content and position of the scene text are generated, and a semantic constellation graph is constructed. This multivehicle system allocates descriptors and achieves precise place recognition through constellation graph matching. Experimental evaluations on the SYNTHIA (SYNTHIA-SEQS-05), VIODE (parking lot), and text-oriented datasets demonstrate that the proposed algorithm attains superior accuracy and robustness, significantly enhancing localization capabilities.