Adaptive high-precision map generation method and system based on context and uncertainty
By modeling attention weights using Dirichlet distribution and adaptively gating computational resources, the robustness and efficiency issues of high-precision map generation methods in complex scenarios are solved, enabling the construction of high-precision autonomous driving maps.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2025-12-15
- Publication Date
- 2026-04-24
AI Technical Summary
Existing high-precision map generation methods waste computational resources and lack robustness in complex scenarios, failing to effectively utilize global context information, resulting in incorrect predictions and low computational efficiency.
We employ Dirichlet distribution to model attention weights and adaptive gating computational resources. Through adaptive point weight fusion and context-aware uncertainty refinement, we dynamically adjust the inference strategy to improve robustness and computational efficiency.
It significantly improves the robustness and computational efficiency of map building in complex scenarios, effectively identifies unseen urban environments, reduces the waste of computing resources, and improves the safety and accuracy of autonomous driving.
Smart Images
Figure CN121921407A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of Internet of Things and intelligent transportation technology, specifically relating to a high-precision map construction technology for autonomous driving scenarios, and in particular a high-precision map generation method and system based on context and uncertainty adaptation. Background Technology
[0002] High-definition maps (HDMaps), serving as the "digital infrastructure" of autonomous driving systems, provide vehicles with beyond-line-of-sight information on road geometry, topology, and traffic rules, providing crucial support for achieving safe navigation in Level 4 / Level 5 autonomous driving. Traditional mapping methods primarily rely on offline scanning and manual annotation using data collection vehicles equipped with expensive LiDAR (Light Detection and Ranging) systems. This approach is not only costly but also has a long update cycle, making it difficult to adapt to the rapid changes in urban road environments.
[0003] In recent years, online vectorized high-precision map construction technology based on vehicle-mounted surround-view cameras has gradually become a research hotspot. For example, methods such as MapTR and VectorMapNet directly regress vector elements such as lane lines and road edges from images through end-to-end deep neural networks. Although these methods have made significant progress on standard datasets, their reliability still faces huge challenges when facing complex open-world scenarios (such as sensor noise caused by severe weather, vehicle occlusion, and blurred intersection geometry).
[0004] The main drawback of existing technologies lies in their use of a "static inference" mechanism, which applies a fixed computational process to all map queries indiscriminately. This mechanism leads to two core problems: first, over-refinement, where the model still performs complex contextual inference in visually clear and high-confidence simple areas, wasting computational resources; second, insufficient inference, where the lack of targeted enhancement mechanisms in occluded or blurred difficult areas prevents the model from effectively utilizing global contextual information to resolve ambiguities, resulting in high-confidence but geometrically incorrect predictions (i.e., the "illusion" phenomenon). Furthermore, existing attention mechanisms often use Softmax for normalization, forcing a deterministic weight distribution that masks the multimodal cognitive uncertainty in feature matching; cross-view fusion strategies are also often based on relatively competitive weight allocation, failing to effectively suppress noisy observation data.
[0005] Therefore, there is an urgent need for an adaptive cartographic framework that can perceive the uncertainty of its own predictions in real time and dynamically adjust its inference strategy accordingly, so as to significantly improve the robustness and security of map construction in complex scenarios while ensuring computational efficiency. Summary of the Invention
[0006] To address the problems mentioned in the background, this invention proposes a high-precision map generation method and system based on context-and-uncertainty-adaptive mapping (CUAM). The core concept of this invention lies in transforming uncertainty from a simple output indicator into a structured control signal guiding network reasoning. It uses Dirichlet distribution to model attention weights to capture cognitive uncertainty and leverages random uncertainty to dynamically gate computational resources, achieving "on-demand reasoning."
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A high-precision map generation method based on context and uncertainty adaptation includes:
[0009] Feature extraction and view transformation stage: Features of the panoramic view are extracted using a shared convolutional neural network backbone, and bird's-eye view (BEV) features are generated through a transformation module.
[0010] Perspective structure prior injection stage: Parallel perspective (PV) branches are designed to directly predict map key points and their pixel-level uncertainties on the image plane. An adaptive point weight fusion (ADFW) mechanism is introduced to calculate fusion weights based on absolute uncertainty, projecting reliable PV structure priors and injecting them into BEV features, effectively suppressing noise interference.
[0011] Dirichlet Uncertainty Attention (DUA) Stage: In the map decoder, the traditional Softmax attention mechanism is abandoned, and the DUA mechanism is proposed. This mechanism predicts the concentration parameters of the Dirichlet distribution and samples the attention weights accordingly. This design not only follows the simplex geometric constraints, but also explicitly quantifies the cognitive uncertainty in the feature correspondence process through the sum of concentration parameters.
[0012] Context-Aware Uncertainty Refinement (CUR) Stage: Constructing a refinement head with dynamic routing capabilities. This module first predicts the base location and uncertainty, then generates a soft-gated signal based on the uncertainty level. For queries with high uncertainty (fuzzy), an additional multi-head self-attention subnetwork is activated, utilizing global contextual information (such as adjacent lane lines and roadside direction) for feature correction; for queries with low uncertainty (clear), this subnetwork is skipped and the output is directly obtained, achieving a dual optimization of computational efficiency and accuracy.
[0013] Furthermore, the method introduces a loss function that includes a negative log-likelihood (NLL) term, co-supervising coordinate regression and uncertainty estimation, forcing the model to reach a consensus between prediction error and output confidence.
[0014] The present invention also provides a cloud-edge collaborative campus abnormal behavior detection system for drones, including a perception and acquisition module, a backbone computing module, a dual-stream processing module, an adaptive decoding module, and a post-processing module, wherein each module works in concert to execute the above method.
[0015] Feature extraction steps: Acquire vehicle surround view image data, and extract multi-view image features through a shared backbone network;
[0016] View transformation and feature fusion steps: The multi-view image features are converted into bird's-eye view BEV features; at the same time, the key points of map elements in the image space and their perceptual uncertainty are predicted by the perspective view PV branch; the PV features are projected and fused into the BEV features using an adaptive point weight fusion strategy.
[0017] Uncertainty-aware decoding steps: Initialize a set of learnable map query vectors, and iteratively regress map elements through cascaded decoder layers; the decoder layer includes a Dirichlet Uncertainty Attention (DUA) module and a Context-Aware Uncertainty Refinement (CUR) module;
[0018] The Dirichlet uncertainty attention module is used to interact map queries with BEV features, generate attention weights by predicting the concentration parameters of the Dirichlet distribution, and calculate the cognitive uncertainty of the query based on the sum of the concentration parameters.
[0019] The context-aware uncertainty refinement module is used to adaptively correct the prediction results of map queries, generate dynamic gating masks based on the random uncertainty of the query, and selectively activate context reasoning paths to optimize the geometric position and uncertainty estimation of fuzzy queries.
[0020] Finally, the output is a vectorized high-precision map containing map element categories, point sequence coordinates, and confidence levels.
[0021] Furthermore, the specific processing steps of the Dirichlet uncertainty attention module include:
[0022] For a given map query, predict Each sampling offset is used to locate key positions on the BEV feature map, while simultaneously predicting a concentration parameter vector. ;
[0023] The concentration parameter vector is generated through a linear transformation and the Softplus activation function, as shown in the formula: ,in For query vector, It is a numerically stable term;
[0024] Based on the concentration parameter vector, from the Dirichlet distribution Attention weight vector obtained by mid-sampling And use this weight to perform weighted aggregation of features at the sampling location;
[0025] Calculate the cognitive uncertainty of the query Defined as the total concentration parameter .
[0026] Furthermore, the context-aware uncertainty refinement module adopts a dual-path reasoning architecture, specifically including:
[0027] Fast path: Performs a linear projection on the input query vector to predict the mean coordinates and underlying uncertainty. ;
[0028] Gating decision: Calculate the average uncertainty of the query and generate a soft route mask using a learnable threshold function. The formula is This is the sharpness coefficient;
[0029] Slow path: Activated when the soft routing mask indicates high uncertainty, it embeds the basic uncertainty into the latent space, captures the global contextual relationships between queries through a multi-head self-attention mechanism, and generates a correction bias. ;
[0030] Fusion Output: The corrected bias is adaptively injected into the base prediction based on the soft routing mask, resulting in a refined uncertainty. ,in This is the scaling factor.
[0031] Furthermore, the adaptive point weight fusion strategy specifically includes:
[0032] In the perspective branch, predict the location of key points in the image coordinate system. and the corresponding pixel-level uncertainty ;
[0033] Using inverse perspective projection to highlight key points Coordinates obtained by elevating to BEV space ;
[0034] Calculate fusion weights based on absolute uncertainty The formula is ,in Temperature coefficient;
[0035] Projected coordinates With uncertainty The joint encoding is converted into a structured cue vector, and the fusion weights are utilized. After modulation, BEV features are injected.
[0036] Furthermore, the training loss function of the method includes a negative log-likelihood loss term (NLL) to supervise the learning of accidental uncertainty;
[0037] The negative log-likelihood loss term is defined as follows: ;
[0038] in, Due to the uncertainty of prediction, To predict coordinates, The coordinates are the true values; by minimizing this loss, the model is forced to increase uncertainty when the prediction error is large, and to reduce uncertainty when the prediction is accurate.
[0039] A high-precision map generation system based on context and uncertainty adaptation includes:
[0040] The perception and acquisition module, configured on the vehicle, is used to collect multi-channel surround view video stream data around the vehicle;
[0041] The backbone computing module is used to receive video stream data and extract multi-scale depth features;
[0042] The dual-stream processing module includes a parallel BEV branch and a PV branch; the BEV branch is used to construct the bird's-eye view feature space, and the PV branch is used to extract the lane structure and uncertainty in the perspective view, and the PV information is injected into the BEV feature via an adaptive fusion unit.
[0043] The adaptive decoding module contains stacked decoding layers, each layer integrating a Dirichlet uncertainty attention unit and a context-aware refinement unit, used to dynamically adjust the inference calculation path according to the fuzziness of the features and output vectorized map elements.
[0044] The post-processing module is used to perform coordinate transformation and topology construction on the decoded vector data to generate the final high-precision map data.
[0045] Furthermore, the Dirichlet uncertainty attention unit in the adaptive decoding module is configured to: abandon deterministic Softmax normalization and probabilistically model the attention weights by predicting the Dirichlet distribution parameters, thereby preserving and transmitting cognitive uncertainty signals during the feature aggregation stage.
[0046] Furthermore, the context-aware refinement unit in the adaptive decoding module is configured to: have a dynamic computation graph function; for queries determined to be of low uncertainty, it outputs directly through only a linear layer; for queries determined to be of high uncertainty, it automatically routes to a sub-network containing multi-head self-attention for global context reasoning.
[0047] Furthermore, the system also includes a temperature modulation module, which adjusts the sensitivity of uncertainty to the fusion weights through learnable temperature parameters during the fusion of the PV branch and the BEV branch, in order to adapt to sensor noise levels under different environments.
[0048] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 5.
[0049] Compared with the prior art, the present invention has the following significant advantages:
[0050] Strong robustness and generalization ability: By explicitly modeling cognitive uncertainty through the DUA module, the model can effectively identify and adapt to unseen urban scenarios (OOD generalization). Experiments show that, under geographically disjoint data partitioning, the average accuracy (mAP) of this invention in cross-city tests is significantly better than existing state-of-the-art methods.
[0051] High computational efficiency: The adaptive routing mechanism of the CUR module avoids redundant computation on simple samples. Experimental data shows that this mechanism improves inference speed by approximately 16% while maintaining accuracy comparable to full inference, and significantly reduces FLOPs, making it more suitable for vehicle-mounted deployment.
[0052] Strong anti-interference capability: The ADFW module performs cross-view fusion based on absolute uncertainty, ensuring that only visually clear and reliable features will affect the BEV representation, effectively filtering out perception noise caused by occlusion or long distance.
[0053] Good geometric consistency: DUA uses the Dirichlet distribution to model attention weights and strictly adheres to the probabilistic simplex constraint, which theoretically guarantees the mathematical rigor of the attention mechanism and can more accurately capture the spatial correspondence of multimodalities. Attached Figure Description
[0054] Figure 1 This is a schematic diagram of the overall architecture of the context- and uncertainty-adaptive high-precision map generation system (CUAM) proposed in this invention; it shows the complete process from image input, BEV / PV dual-stream processing to map decoder.
[0055] Figure 2 This is a schematic diagram of the Dirichlet Uncertainty Attention (DUA) module, illustrating the processes of concentration parameter prediction, Dirichlet resampling, and feature aggregation.
[0056] Figure 3 This is a schematic diagram of the Context-Aware Uncertainty Refinement (CUR) module; it illustrates fast paths, slow paths, and uncertainty-based adaptive routing mechanisms.
[0057] Figure 4 This is a schematic diagram of the Adaptive Point Weighted Fusion (ADFW) module; it illustrates the projection of PV features, uncertainty weighting, and the injection process into BEV features.
[0058] Figure 5 This chart compares the efficiency and accuracy of different routing strategies. Detailed Implementation
[0059] To further illustrate the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings.
[0060] Example 1: System Overall Architecture
[0061] like Figure 1 As shown in this embodiment, a high-precision map generation system based on context and uncertainty adaptation mainly consists of four parts: image backbone network, BEV branch, PV branch, and map decoder.
[0062] Input and feature extraction: The system receives surround view images from six cameras around the vehicle and extracts multi-layer features through a backbone network such as ResNet-50.
[0063] BEV branch: Using transformation techniques such as LSS (Lift-Splat-Shoot) or BEVFormer, perspective view features are converted into a unified BEV feature space.
[0064] PV branch: Parallel to the BEV branch, it performs lightweight predictions directly on image features, outputting 2D keypoint coordinates and their pixel-level uncertainties. These predictions are then fused into the BEV features via the ADFW module (see Example 4 for details).
[0065] Map decoder: includes The cascaded decoding layers each consist of a self-attention module, a DUA module (see Example 2), and a CUR module (see Example 3). The query vector is passed between layers, progressively refining the location and category of map elements.
[0066] Example 2: Dirichlet Uncertainty Attention (DUA) Mechanism
[0067] like Figure 2 As shown, in the map decoding process, traditional DeformableAttention uses Softmax to calculate weights, which suppresses potential ambiguity. This embodiment uses the DUA module:
[0068] For each map query Network prediction Offset of each sampling point and the corresponding concentration parameters The concentration parameter is obtained through a fully connected layer and the Softplus activation function, ensuring non-negativity.
[0069] Attention weight It is no longer a deterministic value, but rather a value derived from the parameter. Random variables sampled from the Dirichlet distribution: .
[0070] Final characteristics It is a weighted sum of the features of the sampling points: .
[0071] Key point: Cognitive uncertainty is defined as... When the model is confused about the sampling location (such as areas with missing texture), the predicted values will be lower. Small values lead to a divergent distribution and uncertainty. Increase. This. It is not only used as an output, but also cascaded to subsequent layers to guide subsequent integration and refinement.
[0072] Example 3: Context-Aware Uncertainty Refinement (CUR) Header
[0073] like Figure 3 As shown, the CUR module aims to solve the problem of "one-size-fits-all" reasoning.
[0074] FastPath: The query vector is first processed by a shared MLP to predict the mean of the coordinates. and raw standard deviation .
[0075] Gating generation: The system calculates the average uncertainty of the current query. Using formulas Generate soft mask .in It is a preset threshold (e.g., 0.5).
[0076] Adaptive routing: If (Low uncertainty) indicates that the current prediction is reliable, and it is output directly. and Skip the tedious calculations.
[0077] like (High uncertainty) Activate the Slow Path. At this point, the query vector and its uncertainty are fed into a Multi-Head Self-Attention (MHSA) module. This module allows the ambiguous query to "observe" surrounding clear queries (e.g., inferring the location of an ambiguous lane line from a clear curb), generating contextual features. .
[0078] Corrected output: Contextual features A deviation term is generated after linear mapping. The final uncertainty is In this way, computational resources are concentrated on the difficult samples that require the most reasoning.
[0079] Example 4: Adaptive Point Weight Fusion (ADFW)
[0080] like Figure 4 As shown, in order to utilize the rich textures of the image space, this embodiment designs an ADFW.
[0081] Unlike traditional methods that use Softmax (i.e., relative competition) to weight multi-view features, ADFW is based on absolute uncertainty.
[0082] PV branch prediction points and its uncertainty Obtained by inverse projection transformation to BEV space .
[0083] The formula for calculating the fusion weight is: .
[0084] This means that if the uncertainty of a visual detection point is high (e.g., it is obscured by a windshield wiper), its weight... The noise level will approach 0, thus being automatically suppressed during fusion; conversely, clear features will be enhanced. This non-competitive fusion method is more suitable for handling outlier noise in open scenes.
[0085] Example 5: Training and Loss Function
[0086] The training of this system employs a bipartite matching strategy. Loss function. It consists of classification loss, regression loss, and uncertainty loss.
[0087] To improve the accuracy of uncertainty estimation, the regression part employs Laplace negative log-likelihood loss (NLLLoss):
[0088]
[0089] This loss function has an adversarial nature: in order to reduce the second term, the model tends to increase... (Acknowledging uncertainty); however, in order to reduce the first term, the model must be reduced. The final equilibrium point forces the model output to accurately reflect the uncertainty of the prediction error.
[0090] This invention constructs an intelligent mapping system that can "know what it doesn't know" and "think deeply" based on that knowledge, just like a human, by organically combining three core components: DUA, CUR, and ADFW. This effectively solves the reliability problem of generating high-precision maps for autonomous driving in complex scenarios.
Claims
1. A high-precision map generation method based on context and uncertainty adaptation, characterized in that, Includes the following steps: Acquire vehicle surround view image data and extract multi-view image features through a shared backbone network; The multi-view image features are converted into bird's-eye view BEV features; at the same time, the key points of map elements in the image space and their perceptual uncertainty are predicted by the perspective view PV branch; and the PV features are projected and fused into the BEV features using an adaptive point weight fusion strategy. A set of learnable map query vectors is initialized, and map elements are iteratively regressed through cascaded decoder layers; the decoder layers include a Dirichlet Uncertainty Attention (DUA) module and a Context-Aware Uncertainty Refinement (CUR) module. The Dirichlet uncertainty attention module is used to interact map queries with BEV features, generate attention weights by predicting the concentration parameters of the Dirichlet distribution, and calculate the cognitive uncertainty of the query based on the sum of the concentration parameters. The context-aware uncertainty refinement module is used to adaptively correct the prediction results of map queries, generate dynamic gating masks based on the random uncertainty of the query, and selectively activate context reasoning paths to optimize the geometric position and uncertainty estimation of fuzzy queries. Finally, the output is a vectorized high-precision map containing map element categories, point sequence coordinates, and confidence levels.
2. The high-precision map generation method based on context and uncertainty adaptation according to claim 1, characterized in that, The specific processing steps of the Dirichlet uncertainty attention module include: For a given map query, predict Each sampling offset To locate key positions on the BEV feature map and simultaneously predict a concentration parameter vector. ; The concentration parameter vector is generated through a linear transformation and the Softplus activation function, as shown in the formula: ,in For query vector, It is a numerically stable term; Based on the concentration parameter vector, from the Dirichlet distribution Attention weight vector obtained by mid-sampling And use this weight to perform weighted aggregation of features at the sampling location; Calculate the cognitive uncertainty of the query Defined as the total concentration parameter .
3. The high-precision map generation method based on context and uncertainty adaptation according to claim 1, characterized in that, The context-aware uncertainty refinement module adopts a dual-path reasoning architecture, specifically including: Fast path: Performs a linear projection on the input query vector to predict the mean coordinates and underlying uncertainty. ; Gating decision: Calculate the average uncertainty of the query and generate a soft route mask using a learnable threshold function. The formula is This is the sharpness coefficient; Slow path: Activated when the soft routing mask indicates high uncertainty, it embeds the underlying uncertainty into the latent space, captures the global contextual relationships between queries through a multi-head self-attention mechanism, and generates a correction bias. ; Fusion Output: The corrected bias is adaptively injected into the base prediction based on the soft routing mask, resulting in a refined uncertainty. ,in This is the scaling factor.
4. The high-precision map generation method based on context and uncertainty adaptation according to claim 1, characterized in that, The adaptive point weight fusion strategy specifically includes: In the perspective branch, predict the location of key points in the image coordinate system. and the corresponding pixel-level uncertainty ; Using inverse perspective projection to highlight key points Coordinates obtained by elevating to BEV space ; Calculate fusion weights based on absolute uncertainty The formula is ,in Temperature coefficient; Projected coordinates With uncertainty The joint encoding is converted into a structured cue vector, and the fusion weights are utilized. After modulation, BEV features are injected.
5. The high-precision map generation method based on context and uncertainty adaptation according to claim 1, characterized in that, The training loss function of the method includes a negative log-likelihood loss term (NLL) to supervise the learning of accidental uncertainty. The negative log-likelihood loss term is defined as follows: ; in, Due to the uncertainty of prediction, To predict coordinates, The coordinates are the true values; by minimizing this loss, the model increases uncertainty when the prediction error is large, and decreases uncertainty when the prediction is accurate.
6. A high-precision map generation system based on context and uncertainty adaptation that implements the method as described in any one of claims 1-5, characterized in that, include: The perception and acquisition module, configured on the vehicle, is used to collect multi-channel surround view video stream data around the vehicle; The backbone computing module is used to receive video stream data and extract multi-scale depth features; The dual-stream processing module includes a parallel BEV branch and a PV branch; the BEV branch is used to construct the bird's-eye view feature space, and the PV branch is used to extract the lane structure and uncertainty under the perspective view, and the PV information is injected into the BEV feature via an adaptive fusion unit. The adaptive decoding module contains stacked decoding layers, each layer integrating a Dirichlet uncertainty attention unit and a context-aware refinement unit, used to dynamically adjust the inference calculation path according to the fuzziness of the features and output vectorized map elements. The post-processing module is used to perform coordinate transformation and topology construction on the decoded vector data to generate the final high-precision map data.
7. The high-precision map generation system based on context and uncertainty adaptation according to claim 6, characterized in that, The Dirichlet uncertainty attention unit in the adaptive decoding module is configured to: abandon deterministic Softmax normalization and probabilistically model the attention weights by predicting the Dirichlet distribution parameters, thereby preserving and transmitting cognitive uncertainty signals during the feature aggregation stage.
8. The high-precision map generation system based on context and uncertainty adaptation according to claim 6, characterized in that, The context-aware refinement unit in the adaptive decoding module is configured to: have dynamic computation graph functionality; for queries determined to be of low uncertainty, output directly through only a linear layer; for queries determined to be of high uncertainty, automatically route to a sub-network containing multi-head self-attention for global context reasoning.
9. The high-precision map generation system based on context and uncertainty adaptation according to claim 6, characterized in that, The system also includes a temperature modulation module, which adjusts the sensitivity of uncertainty to the fusion weights through learnable temperature parameters during the fusion of the PV branch and the BEV branch, in order to adapt to sensor noise levels under different environments.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 5.