Visual place recognition is a fundamental technique for intelligent autonomous navigation systems, especially for loop closure detection and relocalization in simultaneous localization and mapping (SLAM). It aims to determine whether a query image corresponds to a previously visited place by retrieving the most similar image from a geo-tagged database. However, practical visual place recognition remains challenging because scene appearance is often influenced by illumination, weather, seasonal changes, dynamic objects, and perceptual aliasing. Handcrafted descriptors such as local and global visual features are simple and interpretable, but they are usually insufficient for large-scale and long-term navigation in complex environments. Deep learning-based methods have improved scene representation, but they may require considerable computation and memory while also suffering from limited interpretability and feature degradation during network propagation. Semantic information can raise scene understanding from the pixel level to the object level and provide more stable cues under appearance changes. Nevertheless, existing semantic-based approaches still have limitations in integrating contextual relations, boundary information, local geometric structure, and stable multi-layer feature propagation. To address these issues, this paper proposes a visual place recognition model based on multi-criterion extraction and multi-level enhancement of scene semantic information, aiming to improve recognition robustness, retrieval accuracy, and operational efficiency for intelligent autonomous navigation.

The proposed model contains three major modules. First, a multi-criterion semantic feature extraction module was constructed to generate robust semantic scene features. Instead of extracting semantic cues according to a single relationship, this module jointly considered distance, context, and boundary criteria among image pixels or regions. Inspired by multi-criteria decision analysis, it adaptively assigned priority weights and decision thresholds to different criteria. Indifference, preference, and veto thresholds were introduced to describe uncertain semantic relationships, while concordance, discordance, and credibility scores were used to evaluate the joint influence of multiple criteria. This mechanism helped reduce information redundancy, ambiguous object boundaries, and missing local geometric information during semantic feature extraction. Second, a multi-level feature enhancement module was designed to alleviate the loss of shallow semantic information in deep network propagation. The network consisted of four feature-processing stages and three enhancement units. From the second stage onward, shallow semantic features were repeatedly concatenated with deeper features after attention-based feature perception. Through this hierarchical and iterative strategy, the model preserved initial scene information and improved the generalization ability of local features. Third, a rapid spatial scoring module was introduced to improve retrieval efficiency. After matching semantic feature pairs between a query image and candidate images, the module selected semantic feature points in the candidate image and calculated a distance-based spatial score within a local circular region. Candidate images were then ranked according to their semantic spatial scores, enabling the system to obtain the final recognition result rapidly and accurately.

Experiments were conducted on four public datasets, namely Nordland, Pittsburgh, Tokyo 24/7, and RobotCar Seasons v2. These datasets included challenging scenes with illumination variation, weather variation, seasonal change, dynamic object interference, and perceptual aliasing. The proposed model was compared with representative methods including NetVLAD, AP-GEM, SFRS, SuperGlue, DELG, Patch-NetVLAD, TCL, and TransVPR. On Nordland, Pittsburgh, and Tokyo 24/7, the proposed model achieved state-of-the-art Recall@1 values of 65.6%, 92.6%, and 92.5%, respectively. Compared with the best existing results on these three datasets, Recall@1 was improved by 6.8%, 2.1%, and 3.8%, respectively. The model also showed competitive performance on RobotCar Seasons v2 and achieved the best result under the 0.5 m/5° evaluation criterion. Qualitative comparisons showed that the proposed method could correctly recognize places under low illumination, snow occlusion, complex urban road structures, dynamic object interference, and perceptual aliasing, whereas several comparison methods produced incorrect retrieval results in these difficult cases. Robustness was further evaluated by sampling 75%, 50%, and 25% of validation images to simulate incomplete database conditions. When the available image data decreased, the proposed model maintained relatively stable recognition performance, while most comparison methods showed obvious performance degradation. Ablation experiments confirmed the contribution of each module. The multi-criterion semantic feature extraction module improved the baseline by introducing more reliable semantic cues. The multi-level feature enhancement module further improved performance by preserving and repeatedly strengthening shallow semantic information. The rapid spatial scoring module enabled efficient retrieval based on matched semantic features. When all three modules were combined, the model achieved the best overall performance. In terms of efficiency, the proposed model required 42 ms for feature extraction, approximately 1.26 s for feature matching, and only 0.21 MB of memory, indicating a favorable balance between recognition accuracy, running time, and memory consumption.

This paper presents a semantic-enhanced visual place recognition model for intelligent autonomous navigation systems. By jointly modeling distance, context, and boundary criteria, the model extracts more robust semantic features from scenes with complex appearance changes. By hierarchically enhancing shallow features across multiple network levels, it reduces semantic information loss and improves feature representation stability. By introducing a rapid spatial scoring mechanism, it accelerates image retrieval while maintaining reliable recognition accuracy. Experimental results on multiple benchmark datasets demonstrate that the proposed model outperforms or remains highly competitive with existing methods under challenging visual conditions. Its low time and memory requirements make it suitable for real-time or resource-constrained autonomous platforms such as unmanned ground vehicles, unmanned aerial vehicles, and robotic systems used in space-related missions.