Line light sources are chosen as markers due to their balance between information content and practicality. Point sources offer single features that are lost if occluded, while planar markers are complex and unstable underwater. In contrast, lines provide stronger visual robustness: even when partially visible, a single line conveys structural information like direction and length, and multiple lines form rich geometric relations. Thus, using line light sources overcomes the fragility of points while avoiding the high complexity of planes, offering an effective solution for reliable underwater visual location.

Within the effective visual region, the line light source marker is constituted by four line light sources, and four key points are included, as depicted in Figure 2. In the figure, the line light sources are represented by green lines, and points A, B, C, and D are identified as the intersection points of these line light sources (key points) (Xing et al., 2024). Line light source markers located outside the effective field of view are presented in Figure 3. In Figure 3A, a single line light source is observed, and no key points are present. In Figure 3B, two line light sources are included, along with one key point. In Figure 3C, three line light sources are included, accompanied by two key points.

Beyond the effective visual range, incomplete visual information reduces geometric localization accuracy. By analyzing typical patterns of missing information from different directions and classifying them into categories like “left-deviation” or “right-deviation,” preliminary heading alignment can be achieved. This essentially transforms continuous localization into a discrete pattern classification task, reducing reliance on complete environmental perception. The core idea is to leverage statistical correlations between partially visible features and directional deviations, allowing the system to infer approximate orientation through pattern matching even with extensive information loss. This classification-based localization enhances system robustness under weak observation conditions, providing reliable initial constraints for subsequent fine positioning.

x, y, and z represent three observation directions in the local visual field. In underwater docking tasks, priority is typically given to ensuring that the heading of the AUV is aligned with the z-axis direction of the visual marker, with the z-axis of the camera being parallel to the AUV’s heading (Li et al., 2015). Consequently, when considering only the local observation results in the x and y directions, a schematic diagram is presented in Figure 4. In this figure, the boundaries x and y of the effective visual region are extended linearly, thereby dividing the surrounding space into eight directional subspaces. The initial subspace is defined as the region above the effective visual area, and the subspace indices are encoded from 1 to 8 in a clockwise direction. The effective visual region is depicted as the gray area in Figure 4. This region is defined by its bottom-left coordinate (−x_v, −y_v) and top-right coordinate (x_v, y_v). Outside the effective visual region, the area from which line light source information can be observed is bounded by the bottom-left coordinate (−x_n, −y_n) and the top-right coordinate (x_n, y_n). The coordinate regions and their corresponding direction mappings for the subspace are presented in Table 1.

To validate the effectiveness of the proposed model, a local visual localization dataset was constructed in a controlled indoor water tank environment, as illustrated in Figure 8B. The dataset was collected using an AUV equipped with a forward-looking camera, shown in Figure 8A. The visual system captures images at a frame rate of 20 fps, acquiring 20 consecutive frames per second. Simultaneously, marker localization outputs positioning results at a frequency of 1 Hz. As a result, each sequence corresponds to a complete visual positioning cycle, consisting of 20 frames. During the AUV docking process, the typical navigation speed does not exceed 4 knots, approximately 2 m/s. At this speed, the AUV moves about 2 meters within one second, and the spatial sampling interval between adjacent frames captured by the camera is approximately 0.1 meters. Based on the practical requirements of visual positioning, the distance levels are divided into 10 categories. The design of the 20-frame sequence aims to fully utilize the multi-frame information within a single cycle for offline training, enabling the extraction of richer features. Within each subspace, the image acquisition strategy is governed by two dimensions, as expressed in Equation 7.

where x, y, z represents the position coordinates of the AUV’s forward-looking camera. x0,y0 denotes the coordinates of the lower-left corner of the subspace, while x_end, y_end corresponds to the coordinates of the upper-right corner.

For the horizontal direction, a lateral grid sampling method is employed, with the acquisition path defined from left to right and from bottom to top. The fixed acquisition interval Δx, Δy is set to 0.1. For the vertical direction, a longitudinal distance sequence sampling method is used. The initial distance z₀ is set to 1, the cutoff distance z_end is set to 10, and the acquisition interval Δz is set to 1. This process simulates the continuous visual changes observed as the AUV approaches the target from a distance. The subspace number is used as the local visual localization ground truth, and image acquisition within all subspaces is completed sequentially in a clockwise order according to numbers 1–8. Ultimately, the dataset comprises a total of 12,800 image sequences (8 subspaces × 10 distance levels × 160 horizontal sampling points). To assess the model performance, the dataset is randomly partitioned into training, validation, and test sets with a ratio of 6:2:2, as detailed in Table 3. During the division, balanced distribution across subspaces and distance levels is ensured. All data are collected under static, clear water conditions with constant illumination to establish baseline performance.

The model architecture of this paper is structured as follows: Spatial Feature Extraction: The backbone of the CNN primarily consists of two Conv-ReLU-MaxPool modules, which are employed for the initial extraction of spatial features. To further enhance feature representation capability and gradient flow, residual blocks (ResBlocks) are introduced after these two modules, with each residual block comprising two convolutional layers and a skip connection. Finally, the network compresses the feature maps into a fixed-length feature vector through a Global Average Pooling (GAP) layer, replacing the traditional fully connected layers to reduce the number of parameters and improve generalization. The output of GAP can be directly regarded as a flattened feature representation. Spatio-temporal Feature Fusion: The spatial features extracted by the CNN are fed into a spatio-temporal attention Transformer encoder. This encoder employs eight attention heads with a hidden dimension of 64. A learnable temporal decay factor (initialized to 0.95) is introduced to model the relative importance of features at different time steps within the sequence. The parameters selected for the training process are presented in Table 4. All experiments were conducted through offline training on a server equipped with an Nvidia RTX 3090 GPU. The trained model achieved an inference speed of 100 fps on the test set, meeting the real-time requirements. To ensure the statistical reliability of the results and to evaluate their variance, five independent repeated training-testing runs were conducted, each initialized with different random seeds for model weights and data loading order.

The final reported performance is the average of these five runs, yielding a mean accuracy of 98.52%. The narrow confidence interval indicates stable model performance and high reproducibility. The confusion matrix, shown in Figure 9, provides a further analysis of model performance. The results reveal that the classification accuracy across all eight subspaces ranges between 98.1% and 98.8%, demonstrating a relatively balanced distribution without significant performance bias toward any specific subspace (maximum inter-class difference < 5%). This suggests that the model exhibits good and consistent discriminative ability across all subspaces, with no clear signs of overfitting or underfitting observed. The generalization performance is thus validated under the experimental conditions of this study.

The loss and accuracy curves over the first 20 training epochs are presented in Figure 10. The synchronous and rapid descent of both the training and validation loss functions during the initial phases demonstrated efficient model learning and robust generalization capabilities, successfully mitigating significant overfitting. Performance metrics, specifically Top-1 and Top-5 accuracy, exhibited a rapid ascent, with the Top-5 accuracy notably converging quickly toward its theoretical maximum. Crucially, by the 20th epoch, all characteristic curves had plateaued, confirming complete convergence of the training process and the achievement of a stable, steady-state performance level.

To verify the real-time performance of the proposed model in the AUV local visual localization task, systematic comparative experiments were designed and conducted. Two representative baseline models in the field of spatiotemporal sequence processing were selected for comparison: 3D-CNN Model: As an end-to-end spatiotemporal feature extractor, 3D-CNN directly performs three-dimensional convolution operations on the input image sequence (dimensions: time × height × width × channels), aiming to simultaneously capture spatial appearance and temporal dynamic information. LSTM Model: Frame-by-frame images are fed into an LSTM network to model their temporal dependencies. All comparative experiments were conducted under identical dataset splits (training/validation/test sets) and the same training hyperparameters (as shown in Table 4, including optimizer, learning rate, number of epochs, etc.) to ensure a fair comparison. Each compared model was independently trained five times, and its average performance is reported. Detailed comparative results are presented in Table 5.

The proposed model demonstrates significant advantages in classification accuracy, model lightweighting, inference speed, and training efficiency. Its average accuracy reaches 98.52%, which is approximately 2.8 and 1.5 percentage points higher than that of 3D-CNN and LSTM, respectively. The minimal overlap in statistical confidence intervals indicates that the performance improvement is statistically significant. With only 5.2 million parameters, the model is considerably more compact than 3D-CNN (15.3 million) and LSTM (8.7 million). This efficiency is attributed to the effective representational capacity of the Transformer attention mechanism, enabling high performance while maintaining a lightweight architecture suitable for deployment on AUV embedded platforms with limited computational and storage resources. In terms of inference speed, the model achieves 100 fps, significantly outperforming 3D-CNN (42 fps) and LSTM (78 fps). Furthermore, the training process requires only 1.5 hours, demonstrating higher efficiency compared to 3D-CNN (3.8 hours) and highlighting the model’s enhanced overall practicality. To further validate the real-time inference performance of the model on embedded platforms, the models were deployed in the edge computer (Nvidia Jetson Orin NX) carried by the AUV, which offers a computational capacity of 100 TOPS. Considering that the camera mounted on the AUV operates at a frame rate of 20 fps and the system must reserve computational resources for other functions, an image processing capability of 30 fps was set as the real-time safety threshold. In the embedded environment, the model proposed in this paper achieves a stable inference frame rate of 30 fps, meeting the dual requirements of real-time performance and computational headroom. In contrast, the LSTM method only reaches approximately 20 fps, while the 3DCNN method achieves about 13 fps, both of which fail to satisfy real-time processing demands while ensuring sufficient system headroom.

While the algorithm was validated online, an ablation study on the temporal component was conducted, as presented in Table 6, to demonstrate the effectiveness of the spatio-temporal attention mechanism in local visual localization (Hille et al., 2023). The image acquisition frequency is set at 20 fps, with each sequence comprising 20 frames, corresponding to a one-second marker localization cycle. This design ensures both the integrity of the marker localization cycle and the consistency between sequence construction and offline training.

Baseline Model (Comparison Group 1): A static single-frame input architecture was employed, which extracted spatial features exclusively without incorporating temporal information, serving as the fundamental control group. Uniformly Weighted Temporal Model (Comparison Group 2): The input consisted of a sequence of 20 consecutive image frames, and an equal-weight aggregation strategy (Strategy β=120) was adopted. This model assumed equal contribution of features across all time steps, representing the simplest form of temporal integration. Exponentially Decayed Temporal Model (Comparison Group 3): A time decay factor was introduced to construct a temporal architecture (Architecture β_n+1 = β_n/2) with weights decreasing in a geometric progression. In this model, higher weights were assigned to more recent frames, while the weights of subsequent frames decayed according to a half-life pattern.

To validate the universality of the proposed algorithm in three-dimensional space, the basic motion trajectory was first determined in the x–y plane, as illustrated in Figure 11. The AUV maintained uniform motion within each lap. The first, second, and third laps were positioned at the 1/4, 2/4, and 3/4 points between the visual boundary region and the visually effective region, respectively. A total of three laps were executed.

In four independent experiments, the z-axis coordinate was fixed at positions 2, 4, 6, and 8, respectively. In each experiment, motion was initiated from Subspace 1 and proceeded sequentially in a clockwise direction according to the subspace numbering sequence. Unlike image acquisition in offline scenarios, during online validation, the position calculated by the inertial navigation unit of the AUV was used as the ground truth for local visual localization. The results of the local visual localization are presented in Table 7. Accuracy is defined as the percentage of correctly identified marker subspaces relative to the total number of identification attempts.

Compared to the baseline model (Group 1), Group 2, which utilized a uniformly weighted temporal model, was found to suffer from distant-frame interference, resulting in a decrease in accuracy. In contrast, a significant improvement in accuracy was observed for Group 3, where an exponentially decayed temporal model was employed relative to the baseline model (Group 1). In this study, an adaptive temporal decay mechanism was further implemented via a multi-head attention framework to extract spatiotemporal features from multiple dimensions. Experimental results demonstrate that the proposed method achieves superior accuracy compared to the exponentially decayed model (Group 3). Local visual localization is regarded as a continuous process. Following the comparison of accuracy rates, a further comparative analysis was conducted in the temporal domain. A graph was plotted with the subspace index as the horizontal axis and the number of consecutive five recognition errors as the vertical axis.

Curves corresponding to four comparative experimental groups, as illustrated in Figure 12, were generated. Separate curves were plotted for z values of 2, 4, 6, and 8. Additionally, the total number of consecutive five recognition errors was summarized and is presented in Table 8. The baseline model (Group 1), which relied solely on spatial feature extraction, was minimally influenced by temporal continuity. Misidentified events primarily occurred as isolated incidents, with a low frequency of consecutive errors. Although the uniformly weighted temporal model (Group 2) mitigated the occurrence of spurious recognition spikes, an increase in consecutive errors was observed. The exponentially decayed temporal model (Group 3) further suppressed these spurious spikes; however, a higher rate of consecutive errors persisted, indicating that the fusion of spatiotemporal features remained insufficient.

In contrast, the algorithm proposed in this work effectively integrates spatiotemporal features. This integration not only enhances the recognition accuracy of spatial features but also optimizes the inter-frame temporal relationships. A significant reduction in consecutive errors was consequently achieved, validating the superiority of the proposed spatiotemporal joint modeling approach.