Loop-MapNet: A Multi-Modal HDMap Perception Framework with SDMap Dynamic Evolution and Priors

Abstract

High-definition maps (HDMaps) are critical for safe autonomy on structured roads. Yet traditional production—relying on dedicated mapping fleets and manual quality control—is costly and slow, impeding large-scale, frequent updates. Recently, standard-definition maps (SDMaps) derived from remote sensing have been adopted as priors to support HDMap perception, lowering cost but struggling with subtle urban changes and localization drift. We propose Loop-MapNet, a self-evolving, multimodal, closed-loop mapping framework. Loop-MapNet effectively leverages surround-view images, LiDAR point clouds, and SDMaps; it fuses multi-scale vision via a weighted BiFPN, and couples PointPillars BEV and SDMap topology encoders for cross-modal sensing. A Transformer-based bidirectional adaptive cross-attention aligns SDMap with online perception, enabling robust fusion under heterogeneity. We further introduce a confidence-guided masked autoencoder (CG-MAE) that leverages confidence and probabilistic distillation to both capture implicit SDMap priors and enhance the detailed representation of low-confidence HDMap regions. With spatiotemporal consistency checks, Loop-MapNet incrementally updates SDMaps to form a perception–mapping–update loop, compensating remote-sensing latency and enabling online map optimization. On nuScenes, within 120 m, Loop-MapNet attains 61.05% mIoU, surpassing the best baseline by 0.77%. Under extreme localization errors, it maintains 60.46% mIoU, improving robustness by 2.77%; CG-MAE pre-training raises accuracy in low-confidence regions by 1.72%. These results demonstrate advantages in fusion and robustness, moving beyond one-way prior injection and enabling HDMap–SDMap co-evolution for closed-loop autonomy and rapid SDMap refresh from remote sensing.

1. Introduction

HDMaps provide precise geometric and semantic constraints for autonomous driving on structured roads, underpinning robust localization and planning [1,2,3]. However, conventional production pipelines, depending on specialized mapping fleets and intensive manual QA [4,5,6,7], are expensive and slow, limiting scalability and refresh frequency. In parallel, remote sensing (airborne/satellite imagery) enables low-marginal-cost, wide-area coverage with good spatiotemporal accessibility, facilitating the construction of standard-definition maps (SDMaps) [8,9,10] that can act as priors for online perception [11,12,13,14]. SDMaps extend visibility to distant or occluded regions via road topology and lane geometry. Yet, remote sensing latency and localization drift yield geometric errors and semantic ambiguities that degrade prior quality and alignment with onboard sensors.

Two directions dominate online mapping. The first constructs HDMaps purely from onboard sensors [15,16], learning BEV representations from cameras and LiDAR to detect and vectorize map elements; performance nevertheless degrades at distance and in low light. The second injects SDMap priors [13,17] to enhance far-field visibility and mitigate sparsity/occlusions. Existing methods typically assume rigid alignment and are sensitive to GNSS/IMU drift, extrinsic errors, and prior geometry bias; moreover, SDMaps are treated as static priors, lacking closed-loop evolution.

We address the challenge of SDMap-prior-guided HDMap perception with Loop-MapNet: surround-view cameras, LiDAR, and SDMaps are aligned in BEV via a bidirectional adaptive cross-attention; a confidence-guided masked autoencoder (CG-MAE) enhances details in low-confidence regions; and a spatiotemporal-consistency-constrained update yields incremental SDMap refresh, closing the loop of perception–mapping–update–reperception. Compared with P-MapNet [13], HDMapNet [16], and MapTRv2 [18], our design targets real-world remote-sensing and onboard sensors challenges (latency, noise, and cross-view disparities) across three key aspects: (1) Adaptive cross-modal alignment to compensate localization/extrinsic errors and prior geometry bias; (2) Confidence-guided prior enhancement focusing on ambiguous, far, and occluded regions without altering the main inference path; (3) Closed-loop SDMap updates via weighted fusion of high-confidence HDMap predictions and priors under spatiotemporal checks.

On the nuScenes dataset, we validate far-range sensing, misalignment robustness, and update effectiveness. Crucially, the loop significantly improves revisited segments under limited visibility or stale priors, indicating clear engineering value for SDMap–HDMap co-evolution.

2. Related Work

2.1. Online HDMap Construction

Online HDMap construction spans single- and multimodal approaches [19,20,21]. HDMapNet [16] rasterizes BEV from surround cameras via semantic segmentation but suffers from perspective-induced curvature errors at long range. VectorMapNet [21] switches to vectorized map generation with autoregressive Transformer decoding, improving geometric continuity but still facing topological breaks in weak-texture areas. Multimodal methods such as SuperFusion [22] fuse LiDAR and vision to improve accuracy yet remain sensitive to calibration drift; MapTRv2 [18] leverages temporal cues but with limited adaptability. These no-prior approaches are bounded by sensor physics, motivating SDMap priors to extend capabilities.

2.2. SDMap Construction

Remote sensing and crowdsourcing enable low-cost, wide-area SDMap generation [8]. Aerial or satellite pipelines (e.g., RoadTracer [23], Sat2Graph [24], SpaceNet [25]) extract road graphs via CNNs and graph models, yet suffer from breaks and geometric distortions in weak-texture regions [26]. Lightweight semantic maps from crowdsourcing (e.g., RoadMap [27]) build cloud-updated priors for onboard localization at low cost. Despite complementary fusion with SLAM for trajectory correction [28,29], timeliness and fine-grained accuracy remain insufficient for HDMaps; nevertheless, SDMaps are valuable priors for far-field and occluded areas to aid online HDMap construction.

2.3. SDMap-Prior-Aided Mapping

The core is robust alignment and effective fusion between onboard perception and low-accuracy SD priors. P-MapNet [13] integrates SDMap and historical HDMap priors via attention to OpenStreetMap [30], improving lane detection beyond 200 m by 18.73% mIoU. BLOS-BEV [17] fuses near-field perception with SD priors to achieve 0–200 m sensing, gaining over 20% mIoU in 50–200 m. Other works also leverage SD priors for online HDMap construction [12,31]. However, static cross-modal alignment and neglect of localization drift [32] can misalign SD features and onboard observations, leading to prior failures.

Xu et al. [33] generate synthetic deviation maps featuring controlled biases in geometry, topology, and semantics to harden systems against prior discrepancies. However, their approach does not achieve co-evolution between online maps and priors, which consequently allows errors to accumulate across cycles.

2.4. Dynamic Map Update

Dynamic updates include trajectory-based inference, incremental updates, and closed-loop optimization. Trajectory-based approaches such as [34] reconstruct road graphs from GPS traces but falter at complex junctions and sparse data. DeepRoad [35] detects changes via onboard cameras, though binary detection does not yield usable geometries. At city scale, LDMapNet-U [36,36] proposes a perception–update–verification loop to strengthen reliability via multi-source consistency checks. These systems inform our closed-loop SDMap update design.

3. Method

Loop-MapNet aims to construct HDMaps online and update SDMaps in a closed loop through multimodal data fusion and dynamic update mechanisms. As illustrated in Figure 1, Loop-MapNet comprises the following core modules: (1) multimodal feature extraction; (2) cross-modal feature fusion; (3) CG-MAE feature optimization; (4) multi-task prediction and closed-loop update. In this section, we elaborate on the overall architecture and implementation details of each component.

3.1. Feature Extraction

3.1.1. Surround-View Image Feature Extraction

To extract multi-scale features from surround-view cameras and convert them to bird’s-eye view (BEV) representations, this section integrates the core ideas from P-MapNet [13], EfficientDet [37], and Lift, Splat, Shoot [38] to achieve efficient feature extraction and spatial transformation.

Given the surround-view image tensor $I\in {\mathbb{R}}^{B\times n\times 3\times h\times w}$, where B represents the batch size, n is the number of views ($n\in \{1,2,...,6\}$), 3 is the number of RGB channels, and h and w are the height and width of the images, respectively. After merging the batch and view dimensions into $I^{\prime }\in {\mathbb{R}}^{B·n\times 3\times h\times w}$, we process it through the feature extraction network:

$$F_{cam}^{\prime }=\mathrm{CamEncode}\left(I^{\prime }\right)$$

(1)

The output features $F_{cam}^{\prime }\in {\mathbb{R}}^{B·n\times C_{cam}\times h^{\prime }\times w^{\prime }}$, where $C_{cam}$ is the target channel number 256, and $h^{\prime }$ and $w^{\prime }$ are the downsampled dimensions. CamEncode employs EfficientNet-B4 [39] as the backbone network, extracting multi-layer features $\{F_1,F_2,F_3,F_4,F_5\}$, selecting the last three layers of features, unifying the channel numbers through $1\times 1$ convolution, and then fusing through BiFPN:

$$F_{fused}=\mathrm{BiFPN}(C_3^{\prime },C_4^{\prime },C_5)$$

(2)

To adapt to the spatial requirements of BEV transformation, we apply average pooling downsampling to the first layer features in $F_{fused}$, then reshape the dimensions to obtain $F_{cam}\in {\mathbb{R}}^{B\times n\times C_{cam}\times h^{\prime }\times w^{\prime }}$, separating the features for each view.

Based on the Lift, Splat, Shoot [38] approach, we first perform the “Lift” process, flattening the surround-view features into $F_{flat}$, processing the features for each view separately, and lifting 2D features to 3D space through a fully connected network, projecting them onto the BEV plane $F_{bev}$. Subsequently, in the “Splat” process, through inverse perspective mapping (IPM) [40], we project the features from each view to the unified top view $F_{top}$ in the global coordinate system. Finally, we complete the “Shoot” process, optimizing BEV features through upsampling to obtain the BEV feature representation of surround-view images $F_{img}\in {\mathbb{R}}^{B\times C_{img}\times H\times W}$, where B is the batch size, $C_{img}$ is the feature channel number, and H and W are the BEV feature map dimensions.

3.1.2. Point Cloud Feature Extraction

To fully utilize LiDAR’s powerful modeling capability for 3D space, this section adopts PointPillar’s [41] feature extraction method to extract point cloud features and convert them to BEV representation. Given the input point cloud tensor $P_{car}\in {\mathbb{R}}^{N\times D}$, where N is the number of point clouds and D represents the feature dimension of each point (spatial coordinates and reflection intensity, $D=4$). We transform the point clouds from the vehicle coordinate system to the global coordinate system P through vehicle pose, then perform pillarization processing on the transformed point clouds:

$$P_{pillar}=\mathrm{Voxelization}\left(P\right)\in {\mathbb{R}}^{K\times T\times D}$$

(3)

where K is the number of non-empty pillars, T is the maximum number of points per pillar, and D is the feature dimension of each point. We process the point cloud features within each pillar through the PointNet [42] structure:

$$F_{point}=\mathrm{MaxPool}\left(\mathrm{MLP}\left(P_{pillar}\right)\right)$$

(4)

Subsequently, we scatter the point features back to a pseudo-image, process the feature map through a 2D CNN backbone, obtain multi-scale features, and generate the final LiDAR BEV features $F_{lidar}\in {\mathbb{R}}^{B\times C_{lidar}\times H\times W}$ through upsampling and feature fusion, where B is the batch size, $C_{lidar}$ is the feature channel number, and H and W are the BEV feature map dimensions.

3.1.3. SDMap Feature Extraction

The original SDMap serving as prior is provided by the open-source map dataset OpenStreetMap (OSM) [30], containing the road network structure of main roads in the corresponding area. We extract corresponding regional data from OSM based on the vehicle’s pose in the global coordinate system to construct a local SDMap dataset. We represent the original SDMap data in image form and process the SDMap data using the three-layer CNN encoding method from P-MapNet [13].

Representing the SDMap data $M_{sd}$ as a tensor, we employ a three-layer convolutional module for SDMap feature extraction and downsampling, obtaining SDMap features $F_{sd}\in {\mathbb{R}}^{B\times C_{sd}\times H\times W}$ adapted to BEV spatial resolution, where B is the batch size, $C_{sd}$ is the feature channel number, and H and W are the BEV feature map dimensions. In practice, this rasterized interface makes the SDMap encoder source-agnostic: priors from OSM or other providers can be ingested without changing the network. When no prior is available (e.g., new districts or rural roads), the pipeline naturally falls back to pure sensor fusion and accumulates priors over revisits.

3.2. Cross-Modal Feature Fusion

3.2.1. Sensor Feature Fusion

To effectively integrate the respective advantages of surround-view cameras and LiDAR, this section fuses the BEV features of both modalities. In the previous feature extraction process, we have already transformed the surround-view camera features $F_{img}\in {\mathbb{R}}^{B\times C_{img}\times H\times W}$ and LiDAR features $F_{lidar}\in {\mathbb{R}}^{B\times C_{lidar}\times H\times W}$ to a unified global coordinate system and the same spatial resolution through calibration parameters and IPM transformation.

The fusion process adopts channel-wise concatenation:

$$F_{car}=\mathrm{Concat}(F_{img},F_{lidar})\in {\mathbb{R}}^{B\times C_{car}\times H\times W}$$

(5)

where H and W are the feature spatial dimensions corresponding to the perception range, and $C_{car}$ is the total number of feature channels after concatenation (sum of $C_{img}$ and $C_{lidar}$). This direct concatenation fusion approach is computationally efficient while preserving the original information from both modalities, fully utilizing the subsequent SDMap bidirectional adaptive cross-attention mechanism to adaptively learn the importance and correlation of different modal features.

3.2.2. Bidirectional Adaptive Cross-Attention Mechanism

This section proposes an SDMap bidirectional adaptive cross-attention mechanism (bidirectional adaptive cross-attention) aimed at solving multimodal heterogeneous data alignment and effective fusion problems. In practical applications, due to GNSS positioning jitter and insufficient SDMap accuracy, the alignment between sensor features and SDMap features is inaccurate. This mechanism dynamically models the relationships between different modalities and adaptively corrects spatial offsets to generate consistent BEV representations, providing a solid foundation for HDMap online construction and SDMap dynamic optimization.

The inputs include sensor fusion features $F_{car}\in {\mathbb{R}}^{B\times C_{car}\times H\times W}$ and SDMap features $F_{sd}\in {\mathbb{R}}^{B\times C_{sd}\times H\times W}$. The fusion process is divided into three key stages: initial feature processing, adaptive cross-modal alignment and fusion, and feature integration and optimization.

(1) Initial Feature Processing

To enhance the attention mechanism’s perception capability for spatial positions, we first perform positional encoding on the vehicle sensor fusion features. We process $F_{car}$ through convolutional dimensionality reduction to obtain $F_{car}^{\prime }$, then generate sine positi- onal encoding:

$$P_{pos}=\mathrm{PositionEmbeddingSine}(F_{car}^{\prime },\mathrm{mask})$$

(6)

where mask is an all-zero mask, and $P_{pos}$ is the 2D positional encoding containing sine transformation information of spatial coordinates.

(2) Adaptive Cross-Modal Alignment and Fusion

This stage designs a two-layer adaptive cross-attention layer structure to achieve bidirectional feature enhancement and adaptive alignment.

First, we predict inter-modal correlation weights through the feedforward network SimilarityNet:

$$S=\mathrm{SimilarityNet}\left(\mathrm{Concat}(F_{car},F_{sd})\right)$$

(7)

SimilarityNet consists of two linear transformation layers with ReLU activation in the middle and Sigmoid activation in the output layer, ensuring that the output S represents the inter-modal correlation strength. Then, we predict the spatial offset matrix through OffsetNet:

$$O=\mathrm{OffsetNet}\left(F_{car}\right)$$

(8)

OffsetNet also consists of two linear transformation layers, with the output layer using Tanh activation function to generate a normalized offset O, representing the spatial correction amount. Subsequently, we apply the offset to $F_{sd}$ through grid sampling to obtain the dynamically aligned SDMap features $F_{sd}^{\prime }$:

$$F_{sd}^{\prime }=\mathrm{GridSample}(F_{sd},O)$$

(9)

This step overcomes the spatial alignment deviation caused by positioning system jitter and insufficient SDMap accuracy, achieving adaptive spatial correction based on vehicle perception features.

Meanwhile, we enhance the internal consistency of vehicle sensor features through a self-attention mechanism:

$$F_{car}^{\prime }=\mathrm{SelfAttn}(Q_{car},K_{car},V_{car})$$

(10)

where $Q_{car}=K_{car}=F_{car}+P_{pos}$ are the query and key matrices with added positional encoding, and $V_{car}=F_{car}$ is the value matrix.

Finally, we implement cross-modal cross-attention fusion, using vehicle sensor features as queries and the corrected SDMap features as keys and values:

$$F_{aligned}=\mathrm{MultiheadAttn}(Q_{cross},K_{cross},V_{cross})·S$$

(11)

where $Q_{cross}=F_{car}^{\prime }+P_{pos}$ are the enhanced vehicle sensor features, and $K_{cross}=F_{sd}^{\prime }+P_{pos}$ and $V_{cross}=F_{sd}^{\prime }$ are the corrected SDMap features. Using a multi-head attention mechanism combined with residual connections and layer normalization, the final output is adaptively weighted by correlation weights S to obtain the fused feature representation $F_{aligned}$.

After bidirectional attention enhancement and adaptive alignment, we further optimize the feature representation through a feedforward network:

$$F_{aligned}^{\prime }=\mathrm{FFN}\left(F_{aligned}\right)+F_{aligned}$$

(12)

The feedforward network includes dimension expansion, activation function, dropout, and dimension contraction operations, combined with residual connections to preserve original information. Two layers of adaptive cross-attention layer are cascaded to achieve iterative optimization, and finally, the original resolution is restored through upsampling convolution:

$$F_{final}=\mathrm{ConvUp}\left(F_{aligned}^{\prime }\right)$$

(13)

This mechanism overcomes positioning errors by dynamically predicting and correcting spatial offsets through OffsetNet, while SimilarityNet adaptively evaluates the correlation between different modalities to optimize fusion weights, combining self-attention and cross-attention mechanisms to simultaneously strengthen intra-modal consistency and inter-modal complementarity. This design not only solves the limitations of traditional fixed alignment strategies but also fully exploits the complementary information of multimodal data through adaptive learning, providing high-quality fused feature representations for HDMap construction and SDMap updates in complex scenarios.

3.3. Multi-Task Inference Framework

This section utilizes the fused multimodal features $F_{final}$ to achieve multi-task parallel inference through a BevEncode decoder [43] based on a ResNet-18 [44] backbone network. This framework not only generates high-precision HDMaps but also constructs SDMap prediction branches and confidence evaluation branches, forming a complete map perception system. Among them, the SDMap branch is used to update prior maps; the confidence branch quantifies prediction reliability, providing support for the pre-trained confidence-guided masked autoencoder (CG-MAE) to further enhance the detail expression capability of SDMaps in low-confidence regions.

The input is cross-modal fusion features $F_{final}\in {\mathbb{R}}^{B\times C_{\mathrm{in}}\times H\times W}$, where B is the batch size, $C_{\mathrm{in}}$ is the input channel number, and H and W are the height and width of the features, respectively. The output includes four parts: HDMap semantic segmentation map $M_{HD}^{\ast }\in {\mathbb{R}}^{B\times K\times H\times W}$ (K is the number of semantic categories), instance embedding $E_{HD}^{\ast }\in {\mathbb{R}}^{B\times D_e\times H\times W}$ ($D_e$ is the embedding dimension), confidence estimation $C\in {\mathbb{R}}^{B\times 1\times H\times W}$, and SDMap generation result $M_{SD}^{\ast }\in {\mathbb{R}}^{B\times 1\times H\times W}$. The model architecture includes two main stages: feature extraction and multi-task decoding.

3.3.1. Multi-Task Feature Fusion and Encoding

To extract multi-level semantic and spatial information from the input BEV features, we employ a ResNet-18 backbone network for feature extraction. We pass $F_{final}$ through a $7\times 7$ convolutional layer for initial downsampling to generate feature X. Subsequently, after batch normalization and ReLU activation, we sequentially pass through the three residual blocks of ResNet-18 to obtain features of different depths $X_1$, $X_2$, $X_3$.

To achieve efficient decoding while preserving spatial details, we design a multi-task branch architecture based on skip connections. Specifically, we fuse high-level semantic information from deep features $X_3$ with rich spatial details from shallow features $X_1$. Each branch ensures the independence and synergy of multi-task outputs through shared feature fusion frameworks and task-specific decoding modules. This constructs four task-specific decoding branches:

HDMap Semantic Segmentation ${\mathbf{M}}_{\mathbf{HD}}^{\ast }$: We upsample $X_3$ and $X_1$ by 4x to obtain intermediate features $X_{sem}$; then, through 2x bilinear interpolation and convolutional operations, we output the semantic segmentation result $M_{HD}^{\ast }$. Meanwhile, we construct confidence ground truth $C_{gt}$ as supervision signal:

$$C_{gt}=max(\mathrm{softmax}\left(M_{HD}^{\ast }\right),\dim =1)$$

(14)

HDMap Instance Embedding ${\mathbf{E}}_{\mathbf{HD}}^{\ast }$: To achieve instance-level map perception, we fuse $X_3$ and $X_1$, upsample by 4× and adjust channel numbers to generate intermediate features $X_{embed}$; through 2× bilinear interpolation and convolutional operations, we output the instance embedding vector $E_{HD}^{\ast }$.

HDMap Confidence Estimation $C_{HD}^{\ast }$: By predicting confidence scores for each pixel, we provide reliability assessment for HDMap optimization and SDMap updates. We fuse $X_3$ and $X_1$, upsample by 4× and adjust channel numbers to generate $X_{conf}$; after 2× bilinear interpolation and convolutional processing, we output confidence estimation $C_{HD}^{\ast }$ with range [0,1] through Sigmoid activation function.

SDMap Perception Result ${\mathbf{M}}_{\mathbf{SD}}^{\ast }$: We utilize real-time perception results to optimize prior maps. We fuse $X_3$ and $X_1$, upsample by 4× and adjust channel numbers to obtain $X_{sd}$; after 2× bilinear interpolation and convolutional processing, we output binary SDMap result $M_{SD}^{\ast }$ through Sigmoid activation function.

3.3.2. Multi-Task Optimization

To balance the optimization objectives among multiple tasks, we design a weighted combination loss function:

$$L=w_{seg}{L}_{seg}+w_{\mathrm{embed}}{L}_{\mathrm{embed}}+w_{dist}{L}_{dist}+w_{dir}{L}_{dir}+w_{\mathrm{conf}}{L}_{\mathrm{conf}}+w_{hd2sd}{L}_{hd2sd}+w_{sim}{L}_{sim}$$

(15)

where the weight coefficients w are used to balance task priorities. The loss function includes the following key components:

Semantic Segmentation Loss ${\mathbf{L}}_{\mathrm{seg}}$: We adopt weighted binary cross-entropy, alleviating class imbalance problems by assigning higher weights to positive samples.

Embedding Loss ${\mathbf{L}}_{\mathrm{embed}}$: Based on the discriminative loss framework, we combine variance, distance, and regularization losses to optimize intra-instance feature aggregation and inter-instance feature separation. The specific form is:

$$L_{\mathrm{embed}}=L_{variance}+L_{distance}+L_{regularization}$$

(16)

where $L_{variance}$ controls intra-instance feature aggregation, $L_{distance}$ controls inter-instance feature separation, and $L_{regularization}$ prevents excessive feature dispersion.

Distance Loss ${\mathbf{L}}_{\mathbf{dist}}$: As part of the embedding loss, it is used to control distance constraints between instances.

Direction Loss ${\mathbf{L}}_{\mathbf{dir}}$: Implemented using BCELoss, it learns to predict direction information of road elements, helping understand road connection relationships.

Confidence Loss ${\mathbf{L}}_{\mathrm{conf}}$: We use mean squared error to supervise the consistency between confidence prediction and ground truth, enhancing the model’s perception capability for uncertain regions.

HDMap to SDMap Generation Loss ${\mathbf{L}}_{\mathrm{hd}2\mathrm{sd}}$: We adopt binary cross-entropy to optimize SDMap prediction, supporting dynamic map update requirements.

Similarity Loss ${\mathbf{L}}_{\mathrm{sim}}$: Through cross-modal feature alignment, we enhance the collaborative expression capability of multimodal information. The similarity loss measures the consistency between BEV features and SDMap features through cross-modal feature alignment. This loss utilizes the inter-modal correlation weights S predicted by SimilarityNet in the previous equations:

$$L_{\mathrm{sim}}=\alpha ·{\parallel S-S_{gt}\parallel }_2^2+\beta ·\mathrm{CE}(F_{aligned},F_{car})$$

(17)

where $S_{gt}$ is the ground truth correlation indicator calculated based on the feature similarity matrix, obtained by normalizing the cosine similarity of spatially corresponding regions between vehicle perception features $F_{car}$ and SDMap features $F_{sd}$. $\alpha$ and $\beta$ are balance weights. $\alpha$ supervises the accuracy of SimilarityNet prediction, while $\beta$ ensures that the aligned features $F_{aligned}$ retain key information from vehicle perception features $F_{car}$ through cross-entropy loss. This design enables the model to adaptively learn spatial correspondence relationships between different modalities while promoting effective information fusion, providing clear optimization objectives for the aforementioned SDMap bidirectional adaptive cross-attention mechanism.

3.4. Confidence-Guided Masked Autoencoder (CG-MAE)

To address the problem of poor quality in low-confidence regions during HDMap generation, this paper proposes a confidence-guided masked autoencoder (Confidence-Guided Masked Autoencoder, CG-MAE). This method enhances the model’s detail reconstruction capability in low-confidence regions by explicitly modeling the spatial distribution characteristics of HDMaps, thereby effectively improving generation robustness in complex scenarios. Compared with traditional masked autoencoders (MAE) [13] that adopt random or fixed masking strategies, CG-MAE can better handle local uncertainty issues in HDMap generation.

CG-MAE is based on a Vision Transformer architecture, combining feature embedding and multi-task decoding modules, including three stages: feature embedding and masking generation, Transformer encoding and feature reconstruction, and multi-task decoding and optimization. From a system-level perspective, CG-MAE functions as a plug-in refinement pathway that is tightly coupled with the main HDMap branch yet does not alter its inference stability. It first derives dynamic masks from the confidence map $C_{HD}^{\ast }$ using a threshold $\epsilon$ and complements them with a global masking ratio $\theta$ to emphasize uncertain spatial regions while preserving well-determined areas. The masked map is then partitioned and embedded by ConvEmbed, and a ViT-based Transformer encodes long-range structural dependencies of road topology under positional embeddings. A lightweight demasking decoder (MAEHead) reconstructs fine-grained details specifically in masked (low-confidence) regions, producing $O_{opt}$, which is subsequently fused with the primary prediction $M_{HD}^{\ast }$ through confidence gating to yield $O_{fuse}$.

The core idea is shown in Figure 2. By introducing a confidence-guided dynamic masking generation mechanism that prioritizes low-confidence regions, the masked features are fed into the Transformer encoder through positional embedding to capture inter-structural dependencies, and finally, the complete HDMap is reconstructed and detailed expression is enhanced through the decoding module, thereby solving the problem of insufficient details in low-confidence regions during HDMap generation.

This workflow preserves the original multimodal fusion benefits of Loop-MapNet, while explicitly routing representation capacity to ambiguous, far, or occluded areas, thereby enhancing detail fidelity without sacrificing global consistency or introducing optimization instability.

3.4.1. Feature Embedding and Confidence-Guided Masking Generation

The input masked map $M_{\mathrm{masked}}$ is generated through dataset preprocessing, obtained through HDMap confidence estimation $C_{HD}^{\ast }$ (as shown in Figure 3) and random masking operations.

The masking generation process combines confidence distribution $C_{HD}^{\ast }$ with random masking strategy, adopting a two-stage mechanism: first, regions with confidence below threshold $\epsilon$ are prioritized for masking according to global masking ratio $\theta$, then random masking is supplemented throughout the entire image to ensure balanced masking distribution, generating $M_{\mathrm{masked}}$. The feature embedding module divides $M_{\mathrm{masked}}$ into $s\times s$ patches and embeds them into high-dimensional space:

$$X_{\mathrm{embed}}=\mathrm{ConvEmbed}\left(M_{\mathrm{masked}}\right)$$

(18)

where ConvEmbed employs $s\times s$ convolutional kernels and stride s convolutional operations, outputting $X_{\mathrm{embed}}\in {\mathbb{R}}^{B\times D\times H\times W}$.

3.4.2. Transformer Encoding and Feature Reconstruction

Following the idea of Vision Transformer (ViT) [45], the masked feature sequences are processed through positional embedding and multi-layer Transformer:

$$X_{enc}=\mathrm{Transformer}(X_{\mathrm{embed}}+P_{pos})$$

(19)

where $P_{pos}$ is the positional encoding consistent with the previous Section 3.2.2, Transformer contains 12 layers of attention mechanisms, outputting $X_{enc}\in {\mathbb{R}}^{B\times N\times D}$, where $N=H\times W$ is the total number of patches. The decoding stage restores spatial resolution through two layers of deconvolution with 2x upsampling and bilinear interpolation:

$$X_{recon}=\mathrm{UpSample}\left(X_{enc}\right)$$

(20)

obtaining $X_{recon}\in {\mathbb{R}}^{B\times D\times H\times W}$ for reconstructing the original features of masked regions.

3.4.3. Decoding, Optimization and Training Strategy

The reconstructed features $X_{recon}$ are processed through the decoding module (MAEHead, containing only demasking decoding and segmentation head, not including the aforementioned Transformer encoder) to generate the optimized HDMap output $O_{\mathrm{opt}}$:

$$O_{\mathrm{opt}}=\mathrm{MAEHead}\left(X_{recon}\right)$$

(21)

CG-MAE training follows three stages:

(1) Backbone Warm-up: Train the Loop-MapNet backbone network (3.1–3.3) without connecting CG-MAE to obtain stable $M_{HD}^{\ast }$ and $C_{HD}^{\ast }$. $C_{HD}^{\ast }$ is used to offline generate confidence-guided masking samples;

(2) CG-MAE Pre-training: Freeze the backbone and train CG-MAE using the above samples to prioritize reconstruction of low-confidence regions;

(3) Integration and Fine-tuning: Connect the pre-trained CG-MAE to the end of the main model to generate $O_{opt}$, and fuse it with $M_{HD}^{\ast }$ through confidence gating to obtain $O_{fuse}$. During joint fine-tuning, the backbone maintains the original multi-task loss, and by default freezes the CG-MAE encoder, only fine-tuning the MAEHead and upsampling layers.

3.5. Dynamic SDMap Update Strategy

Reviewing the design motivation of this module, Loop-MapNet aims to enhance HDMap real-time perception using low-cost SDMap priors. To address the problem of HDMap accuracy degradation caused by SDMap static errors and lack of global consistency in online inference, we implement consistency enhancement and iterative optimization through a dynamic update strategy that fuses HDMap high-confidence regions with SDMap inference differences.

This strategy processes the results from the previous model output $C_{HD}^{\ast }$, $M_{SD}^{\ast }$, and prior SDMap data $M_{SD}$ through Sigmoid activation to obtain probability maps, achieving dynamic SDMap updates through confidence weighting, updating the SDMap database while gradually improving SDMap accuracy and its enhancement effect on HDMap generation.

First, we extract high-confidence region masks from the confidence map $C_{HD}^{\ast }$ using the confidence threshold $\theta$ to screen reliable regions:

$$M_{\mathrm{conf}}=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}C>\theta \hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(22)

Then, we calculate the difference between the inferred SDMap $M_{SD}^{\ast }$ and the original SDMap $M_{SD}$:

$$\Delta M=(M_{SD}^{\ast }-M_{SD})\in [-1,1]$$

(23)

According to confidence $C_{HD}^{\ast }$ and mask $M_{\mathrm{conf}}$, we weight and fuse the differences to focus on credible region updates:

$$M_{\mathrm{sd}\_\mathrm{new}}=M_{\mathrm{sd}\_\mathrm{ori}}+\alpha ·C_{HD}^{\ast }·M_{\mathrm{conf}}·\Delta M$$

(24)

where $\alpha \in (0,1)$ is the update step size, ensuring that high-confidence regions dominate the update through $C_{HD}^{\ast }·M_{\mathrm{conf}}$.

Subsequently, $M_{\mathrm{sd}\_\mathrm{new}}$ is constrained to $[0,1]$ through Sigmoid:

$$M_{\mathrm{sd}\_\mathrm{new}}=\mathrm{Sigmoid}\left(M_{\mathrm{sd}\_\mathrm{new}}\right)$$

(25)

Finally, we use the updated $M_{\mathrm{sd}\_\mathrm{new}}$ to replace $M_{SD}$ as the prior for the next round of inference, forming a closed-loop optimization to improve the robustness and accuracy of HDMap generation.

4. Experiments and Discussion

4.1. Implementation Details

Dataset. Our experiments utilize the nuScenes dataset, comprising 1000 scenes with each scene lasting 20 s at a sampling frequency of 2 Hz, covering diverse urban environments in Boston and Singapore. The dataset provides six surround-view camera images (1600 × 900 resolution), 32-beam LiDAR point clouds, and high-precision HDMap annotations. We follow the official training set (700 scenes) and validation set (150 scenes) division, with the test set specifically selected to include typical scenarios such as crossroads, T-junctions, and curves. OpenStreetMap (OSM) provides SDMap priors, which are preprocessed and aligned with the nuScenes coordinate system to simulate low-cost map information.

Evaluation Metrics. We employ mean Intersection over Union (mIoU) for semantic segmentation, evaluating the accuracy of three object categories: Drivable areas (Div), Pedestrian areas (Ped), and road Boundaries (Bound). Additionally, we monitor model inference speed (FPS) to assess real-time performance.

Experimental Settings. The network is trained using the Adam optimizer with an initial learning rate of 0.0005, employing a step decay scheduling strategy (adjusted every 10 epochs). The batch size is set to 8, training for 30 epochs. The multi-task loss function weights are configured as $w_{\mathrm{seg}}=1.0$, $w_{\mathrm{embed}}=0.1$, $w_{dist}=0.1$, $w_{dir}=0.1$, $w_{\mathrm{conf}}=0.1$, $w_{hd2sd}=0.5$, $w_{sim}=0.1$. The balance weights for similarity loss $L_{\mathrm{sim}}$ are $\alpha =0.6$ and $\beta =0.4$. In CG-MAE pre-training, the confidence threshold $\epsilon$ is set to 0.6, and the global masking ratio $\theta$ is 0.5. The image input size is 128 × 352, and point cloud ranges are configured with three different scales of perception ranges according to experimental requirements: 60 × 30 m, 120 × 60 m, and 240 × 60 m, to evaluate model performance under different scenarios.

Experimental Environment. Implementation is based on PyTorch 1.10, using 3×NVIDIA Tesla A100 80G for parallel training, with a single NVIDIA GeForce RTX 3090 for inference performance evaluation.

4.2. Far-Range Perception Experiments

To systematically evaluate the far-range perception performance of Loop-MapNet’s multimodal perception enhancement architecture, we configured three different scales of perception ranges: 60 × 30 m, 120 × 60 m, and 240 × 60 m, conducting comparative analysis with HDMapNet [16] and P-MapNet [13].

As shown in

Table 1. Quantitative results across different perception ranges. Best results in bold.

Range	Method	Epoch	Modal	Div.	Ped.	Bound.	mIoU	FPS
60 × 30	HDMapNet [16]	30	C + L	45.13	30.76	56.05	43.98	21.62
	P-MapNet [13]	30	C + L + SD	53.35	39.81	63.36	52.17	9.65
	Ours	30	C + L + SD	54.26	40.13	63.97	52.79	8.10
120 × 60	HDMapNet [16]	30	C+L	54.18	38.03	57.92	50.04	21.55
	P-MapNet [13]	30	C + L + SD	63.64	50.25	66.95	60.28	9.92
	Ours	30	C + L + SD	64.26	51.32	67.58	61.05	7.53
240 × 60	HDMapNet [16]	30	C+L	39.69	26.42	43.53	36.55	13.58
	P-MapNet [13]	30	C + L + SD	52.46	41.83	53.56	49.28	6.92
	Ours	30	C + L + SD	53.26	42.14	54.37	49.92	6.50

, within the 60×30m range, our method achieves an average accuracy of 52.79%, improving by 8.81% and 0.62% compared to HDMapNet and P-MapNet [13], respectively. When the perception range extends to 120 × 60 m, our method reaches 61.05% average accuracy, improving by 9.66% and 0.63% compared to HDMapNet and P-MapNet [13], respectively. When the perception range extends to 240 × 60 m, due to environmental detail degradation at long distances captured by sensors, HDMapNet’s accuracy significantly declines, while Loop-MapNet maintains 49.92% accuracy, validating the effectiveness of multimodal fusion for far-range perception.

As illustrated in Figure 4, within the maximum test range of 240 × 60 m, the actual field of view approaches the physical perception limits of sensors. Figure 4 demonstrates our method’s clear advantages over baseline algorithms in far-range intersection details and lane line position detail expression. Figure 5 shows that under night scenarios, HDMapNet [16] exhibits mapping inaccuracy in far-range areas, while Loop-MapNet can more accurately reconstruct intersection structures and lane lines, providing sufficient prior information for vehicle advance planning.

Although our method’s inference speed is slightly lower than baseline algorithms, in HDMap mapping tasks, mapping accuracy and quality are far more important than real-time requirements, making this performance trade-off acceptable in practical applications.

4.3. Robustness to SDMap Prior Alignment Errors

To evaluate the system’s robustness to alignment errors between SDMap priors and sensor data, we simulate localization uncertainty under urban canyon effects by adding random perturbations to the ego vehicle position and heading in the nuScenes dataset. The experiments apply maximum random perturbations of 5 m and 20 m to the ego vehicle position coordinates, while simultaneously applying 1° and 5° perturbations to heading angles, selecting intersection scenarios for comparative analysis.

As shown in

Table 2. Quantitative results under different localization errors at 120 × 60 m perception range. Best results in bold.

Max Pos.	Max Yaw	Method	Div.	Ped.	Bound.	mIoU
0 m	0°	P-MapNet [13]	63.64	50.25	66.95	60.28
		Ours	64.26	51.32	67.58	61.05
5 m	1°	P-MapNet [13]	63.31	49.87	66.03	59.74
		Ours	64.07	51.43	67.45	60.98
20 m	5°	P-MapNet [13]	60.75	47.63	62.37	56.92
		Ours	63.96	51.05	66.38	60.46

, under conditions without localization perturbations, our method outperforms P-MapNet [13] across all metrics, with mIoU improving by 0.77%. When introducing minor perturbations (5 m/1°), P-MapNet’s mIoU drops by 0.54%, while our method drops by only 0.07%, demonstrating excellent error adaptation capability. Under extreme perturbations (20 m/5°), P-MapNet’s mIoU drops by 3.36%, while our method drops by only 0.59%, fully demonstrating the proposed method’s robustness to significant localization deviations.

As shown in Figure 6, under extreme localization error conditions (20 m/5°), P-MapNet experiences severe SDMap prior misalignment leading to delayed mapping of forward intersections, while Loop-MapNet effectively compensates for localization deviations through the bidirectional adaptive cross-attention mechanism, maintaining accurate mapping performance. Particularly in boundary regions (Bound.), our method drops by only 1.2% under maximum perturbations, while the baseline method drops by 4.58%, validating the effectiveness of the adaptive feature alignment mechanism in maintaining road topology structure.

4.4. Ablation Study: CG-MAE Pre-Training Models

To verify the optimization effect of the confidence-guided masked autoencoder (CG-MAE) on HDMap mapping details, we designed comparative experiments for pre-training models: (1) no-pre-training baseline; (2) traditional MAE pre-training [13,46]; (3) the proposed CG-MAE pre-training. All experiments use the same training cycle (30 epochs) and perception range (120 × 60 m).

As shown in

Table 3. SDMap inference metrics under different pre-training models at 120 × 60 m perception range. Best results in bold.

Pre-Training	Div.	Ped.	Bound.	mIoU	FPS
None	62.67	49.68	65.63	59.33	13.8
MAE [13,46]	63.14	50.26	67.08	60.16	7.36
CG-MAE (Ours)	64.26	51.32	67.58	61.05	6.53

, compared to the no-pre-training baseline, MAE pre-training improves mIoU by 0.83%, while CG-MAE further improves to 61.05%, achieving a 1.72% improvement. Particularly in pedestrian areas (Ped.) and road boundary (Bound.) metrics, CG-MAE improves by 1.64% and 1.95%, respectively, compared to the no pre- training baseline.

As shown in Figure 7, under rainy low-visibility scenarios, the no-pre-training model can perceive distant intersections but exhibits blurred road edges and inaccurate topological connections; MAE pre-training improves intersection structure expression but still contains errors in junction connection relationships; CG-MAE pre-training significantly enhances complex intersection topological details, maintaining clear edge features even under low-visibility conditions.

CG-MAE, through its confidence-guided dynamic masking strategy, enables the model to focus more on data uncertain regions during pre-training, learning more discriminative feature representations. Experiments demonstrate that this strategy, while inheriting the advantages of self-supervised learning, significantly enhances HDMap detail perception capability in low-confidence regions, providing more reliable high-precision map support for autonomous driving systems.

4.5. Impact of HDMap Closed-Loop Updates on SDMap Perception Performance Under Road Network Structure Changes

To evaluate the performance improvement of Loop-MapNet’s closed-loop update mechanism under SDMap update lag scenarios, we simulate SDMap update lag through manual road network structure perturbations, focusing on verifying the enhancement effect of the closed-loop data on HDMap perception accuracy. The experiments specifically select visibility-limited scenarios, such as rainy, foggy weather and nighttime, for testing.

Considering that local road network changes have a limited impact on global IoU metrics, this experiment primarily employs qualitative analysis methods, focusing on examining the improvement effect of closed-loop update mechanisms on SDMap prior quality.

Experimental results indicate that under rainy, foggy weather conditions (Figure 8), the original SDMap prior fails to model distant intersections due to visibility limitations. After Loop-MapNet closed-loop updates, the system achieves accurate intersection reconstruction under the same visibility conditions, with intersection connection relationships and lane edge geometric expressions significantly improved.

Night scenario testing (Figure 9) shows that the original SDMap is limited by camera visibility in narrow road areas, primarily relying on immediate perception from onboard sensors. Through closed-loop updates, when the system revisits the same section, it can accurately reconstruct road topology structures even under equivalent visibility conditions, validating the compensatory effect of closed-loop data on sensor limitations.

By feeding high-confidence HDMap predictions back to SDMap updates, Loop-MapNet effectively improves reconstruction quality at changed intersections, particularly excelling in visibility-limited environments. This closed-loop data mechanism ensures continuous optimization of SDMap priors, providing a more reliable foundation for HDMap construction and offering an effective solution for long-term maintenance of high-precision maps.

5. Conclusions

The Loop-MapNet framework proposed in this paper innovatively realizes a “perception-mapping-update-reperception” closed-loop system, breaking through the traditional one-way prior injection paradigm and achieving collaborative evolution between HDMaps and SDMaps for the first time. The main innovations include: (1) bidirectional adaptive cross-attention mechanism for dynamic alignment of heterogeneous features; (2) confidence-guided masked autoencoder for enhancing detail expression in low-confidence regions; (3) dynamic SDMap update strategy based on high-confidence predictions, forming a closed-loop ecosystem of map data.

Experiments validate the framework’s superior performance across multiple challenging scenarios: the multimodal fusion architecture significantly outperforms single-modal methods in far-range perception; the adaptive alignment mechanism maintains stable performance under extreme localization errors; CG-MAE pre-training significantly improves mapping accuracy in low-visibility regions; the closed-loop update mechanism effectively improves perception capability for changing road networks.

Future work will focus on the following: (1) optimizing computational efficiency for real-time deployment (e.g., pruning/quantization/distillation and TensorRT/edge acceleration toward >15 FPS without accuracy loss); (2) expanding multi-vehicle collaborative updates with confidence-weighted fusion, conflict resolution, and privacy-aware communication to build a collective intelligent map ecosystem; (3) enhancing cross-city generalization and urban-scale scalability via hierarchical map representations and cloud–edge collaboration for efficient storage, retrieval, and incremental updates; (4) establishing a raster-to-vector pipeline and cloud-side global fusion to translate local raster updates into city-scale vector maps.