Method for looking around to occupy grids in combination with on-line vector map feature priori

Abstract

The invention discloses a look-around occupied grid method combined with online vector map feature prior, relates to the technical field of computer vision and automatic driving, and solves the problems that a traditional occupied grid construction method generally depends on data of a laser radar and other sensors, ignores uncertainty features of map information, and is low in cost. In order to solve the problem that in the prior art, the environment understanding ability in a complex scene is insufficient, and particularly limitation is shown in a dynamic traffic scene, the method comprises the following steps that S1, detailed modeling is conducted on uncertainty characteristics of an online map, that is, joint modeling is conducted on uncertainty of positions and categories of map elements, and more stable occupation prediction is provided; and the uncertainty in the vectorized map features is quantized by adopting a mode based on Laplacian distribution. According to the method, the training convergence speed can be remarkably increased, the understanding and decision-making ability of the automatic driving vehicle to the surrounding environment is enhanced, and particularly, higher driving safety and reliability are achieved in a dynamic mixed traffic scene.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and autonomous driving technology, and in particular to a surround view occupancy grid method combined with online vector map feature priors. Background Art

[0002] With the rapid development of autonomous driving technology, high-precision maps and occupancy grid technology are becoming increasingly important in autonomous driving systems. High-precision maps provide static environmental information for autonomous vehicles, including detailed road topology and traffic rules such as lane lines, crosswalks, and curbs. This information is essential for path planning and decision-making. However, traditional high-precision maps are expensive to build, have low timeliness, and are difficult to adapt to the frequently changing environment in open road scenarios.

[0003] In recent years, local semantic maps built from a bird's-eye view have gradually gained favor among researchers due to their efficient computing and lower costs. However, rasterized maps often contain a lot of redundant information, have high memory overhead, high transmission bandwidth requirements, and cannot be directly used for downstream planning and prediction tasks. Therefore, researchers began to explore methods for constructing vectorized maps, which can provide more compact and efficient environmental representation, reduce data redundancy, and improve system real-time performance;

[0004] In this context, occupancy grid (Occ) technology is widely used in autonomous driving environment modeling. The occupancy grid method divides the environment into several small grids and marks whether each grid is occupied, thereby providing real-time dynamic environment information for the autonomous driving system. However, traditional occupancy grid construction methods usually rely on data from lidar and other sensors, ignoring the uncertainty characteristics of map information, resulting in insufficient understanding of the environment in complex scenes, especially in dynamic traffic scenes; therefore, it does not meet existing needs. In this regard, we propose a surround occupancy grid method that combines online vector map feature priors. Summary of the invention

[0005] The purpose of the present invention is to provide a surround occupancy grid method combined with online vector map feature priors to solve the problem that the traditional occupancy grid construction method proposed in the above background technology usually relies on data from lidar and other sensors, ignoring the uncertainty characteristics of map information, resulting in insufficient ability to understand the environment in complex scenes, especially showing limitations in dynamic traffic scenes.

[0006] To achieve the above object, the present invention provides the following technical solution: a look-around occupancy grid method combined with online vector map feature priors, comprising the following steps:

[0007] S1: Detailed modeling of online map uncertainty features, that is, jointly modeling the uncertainty of map element positions and categories to provide more robust occupancy predictions, using a Laplace distribution-based approach to quantify the uncertainty in vectorized map features, and completing the modeling and representation of map uncertainty features;

[0008] S2: Using multi-camera input data, by fusing the multi-scale features of each camera, to achieve more accurate 3D occupancy prediction, and complete the fusion of multi-camera 3D occupancy prediction;

[0009] S3: The uncertainty features in the map are used as prior information and jointly modeled with the occupancy prediction task to complete the joint modeling of map uncertainty features and occupancy prediction.

[0010] Preferably, the map uncertainty feature modeling and representation comprises the following steps:

[0011] D1: Position uncertainty modeling: Online vector map estimation usually uses regression methods to predict the positions of polygon vertices, expressed as position vectors v=v1,v2) in a two-dimensional BEV coordinate system. A regression head containing position and scale parameters is used to convert this prediction result into a probability model. Assume that each map element consists of multiple vertices, and the vertex set is represented as in is the position of the i-th vertex. For each vertex position, the Laplace distribution is used to describe its uncertainty. The distribution formula is:

[0012]

[0013] in, is the expected value of the position, is the scale parameter, which indicates the uncertainty of the jth coordinate of the i-th vertex;

[0014] D2: Constructing a joint probability density function: A map element M containing multiple vertices is used to comprehensively describe the uncertainty of the entire map element by combining the uncertainties of all vertices, so that the overall uncertainty of the online map is better characterized;

[0015] D3: Probabilistic modeling of the output of the category classifier. The classification head will output the confidence score of the category to which each vertex belongs. The confidence score is the probability distribution of the classifier. For each vertex v (i) , let its category label be c (i) ,The uncertainty of the category is quantified by the distribution of classification confidence, so that the entire map element has uncertainty measures in terms of both location and category;

[0016] D4: Design a loss function to replace the traditional mean square error regression loss with negative log likelihood (NLL) loss to capture the uncertainty of the position.

[0017] Preferably, the joint probability density function formula is:

[0018]

[0019] Preferably, the loss function formula is:

[0020]

[0021] This loss function encourages the model to optimize both the expected and uncertainty parameters of the location.

[0022] Preferably, the fusion of multi-camera 3D occupancy prediction comprises the following steps:

[0023] E1: Multi-camera feature extraction and representation: Given a set of multi-camera image inputs {I 1 ,I 2 ,…,I N}, a backbone network is used to extract the features of each camera, and the feature extraction of each camera produces a multi-scale feature representation, denoted as in represents the features of the jjjth scale from the iiith camera;

[0024] E2: Adopting the Transformer-based spatial cross attention mechanism, using the spatial relationship between multiple cameras to weight the fusion features, given the features from different cameras By cross-attention, the features Aggregate to obtain a three-dimensional convolutional volume feature V;

[0025] E3: 2D to 3D spatial mapping: When implementing 3D occupancy prediction, 2D features are fused through a 2D to 3D mapping operation. Using the deformation attention mechanism, 3D voxel reference points are projected onto the 2D view and features around these 2D points are sampled for fusion. For each 3D query point q p , project it into the corresponding 2D view, and the output features It is the result of weighted summation based on the cross-view attention mechanism:

[0026]

[0027] in, represents the set of viewing angles from which the 3D reference point can be observed, Q pis the 3D volume query of the pth element, and DeformAttn represents the weighted feature calculated by deformation attention;

[0028] E4: Multi-scale feature fusion and prediction: We make full use of multi-level feature information and adopt a 2D-3D U-Net architecture to extract and fuse 3D convolutional volume features at multiple scales through 2D to 3D spatial mapping operations.

[0029] E5: Determine the loss function and supervision mechanism: The model can be effectively supervised at different scales, and a supervision signal is introduced at each scale. Specifically, for the occupancy prediction results of each scale, the present invention uses cross entropy loss and affinity loss of scene categories as supervision signals, and the loss function is:

[0030]

[0031] in, is the attenuated loss weight.

[0032] Preferably, the formula of the three-dimensional convolution volume feature V is:

[0033] V=G(I 1 ,I 2 ,…,I N );

[0034] Where G is a neural network module that uses multi-head cross attention to map the 2D features of multiple cameras into 3D occupancy probabilities.

[0035] Preferably, 3D convolutional volume features are extracted and fused at multiple scales to give multi-scale 2D features. After multiple 2D-3D spatial attention layers, multi-scale 3D convolution features are obtained Then the 3D convolutional features of the j-1th layer are upsampled and fused with the features of the jth layer to obtain Y j , the formula is:

[0036] Y j =F j +Deconv(Y j-1 );

[0037] By combining high-level semantic features and low-level detail features, occupancy prediction results with different resolutions are generated.

[0038] Preferably, the joint modeling of map uncertainty features and occupancy prediction comprises the following steps:

[0039] F1: Encode map features: Encode the uncertainty of each map element. Each map element M is composed of several vertices. Let the position of the i-th vertex be v (i) , the uncertainty is given by the location parameter μ (i) and scale parameter b (i) Indicates that the category probability c is introduced (i) , used to describe the category information of map elements, combine the above information, encode the characteristics of each vertex, and the encoding result is expressed as:

[0040]

[0041] in is the uncertainty characteristic of the map;

[0042] F2: Joint modeling and fusion: Combine the occupancy grid feature V obtained by multi-camera fusion, take the uncertainty information of the map as a priori, and jointly perform occupancy prediction. Let the fused occupancy grid feature be V = [v1, v2, …, v M ], each feature v j Corresponding to a voxel in a three-dimensional space, the uncertainty characteristics of the map Fusion into the occupancy prediction of each voxel to enhance the prediction of occupancy state;

[0043] For the occupancy state prediction of each voxel, the following formula is used for joint modeling:

[0044]

[0045] Among them, f() is a joint modeling function, which uses linear weighting or nonlinear functions to fuse the occupancy features from multiple cameras and the uncertainty features of the map;

[0046] F3: Design loss functions to enable the model to effectively learn joint modeling: Establish multi-task loss functions Including cross entropy loss for occupancy states Regularization loss fused with uncertainty features The formula is:

[0047]

[0048] in is the cross entropy loss of the occupied state, is the regularization loss of uncertainty fusion, and λ is the trade-off coefficient.

[0049] Preferably, a linear fusion function is also used in step F2 to weight the map prior and occupancy features, and then the output range is limited by an activation function:

[0050]

[0051] Among them, σ(·) is the activation function, w1 and w2 are learnable parameters used to adjust the influence from different information sources.

[0052] Preferably, the Used to measure the error between the predicted voxel occupancy state and the true annotation, To ensure that the model does not overfit to sensor noise or defective areas when considering map uncertainty information, λ is used to balance the contribution of occupancy prediction error and uncertainty regularization to the overall loss.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The present invention improves the prediction performance and training efficiency of the autonomous driving system in complex road scenarios by combining the environmental information provided by the online vector map. By modeling the uncertainty features extracted from the vector map into the grid occupancy task, the present invention can significantly accelerate the training convergence speed and enhance the autonomous driving vehicle's understanding and decision-making ability of the surrounding environment, especially in dynamic mixed traffic scenarios to achieve higher driving safety and reliability;

[0055] 2. The present invention uses uncertainty features to improve the prediction accuracy of the occupancy grid, especially in the presence of occlusion or sensor noise. The joint modeling function usually uses linear weighting or nonlinear functions to fuse the occupancy features from multiple cameras and the uncertainty features of the map to adjust the influence from different information sources. In this way, the model can combine the uncertainty information of the map in the occupancy state prediction, thereby outputting more conservative estimates in areas with higher uncertainty, improving the system's adaptability to dynamic environments.

[0056] 3. Through joint modeling and loss function design, the present invention can effectively capture environmental uncertainty in complex and dynamic traffic environments, thereby improving the accuracy and robustness of occupancy grid prediction. The main advantage of this method is that it combines the uncertainty of map information, especially in the context of multi-camera data fusion, so that the prediction results can still maintain high accuracy and stability in the presence of occupancy and sensor noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the flow chart of the map uncertainty estimation method of the present invention;

[0058] Figure 2 It is a schematic diagram of the fusion process of multi-camera 3D occupancy prediction of the present invention;

[0059] Figure 3Schematic diagram of the joint modeling process of map uncertainty characteristics and occupancy prediction of the present invention. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0061] See also Figures 1 to 3 , an embodiment of the present invention provides: a look-around occupancy grid method combined with online vector map feature priors, comprising the following steps:

[0062] S1: Detailed modeling of online map uncertainty features, that is, jointly modeling the uncertainty of map element positions and categories to provide more robust occupancy predictions, using a Laplace distribution-based approach to quantify the uncertainty in vectorized map features, and completing the modeling and representation of map uncertainty features;

[0063] S2: Using multi-camera input data, by fusing the multi-scale features of each camera, to achieve more accurate 3D occupancy prediction, and complete the fusion of multi-camera 3D occupancy prediction;

[0064] S3: The uncertainty features in the map are used as prior information and jointly modeled with the occupancy prediction task to complete the joint modeling of map uncertainty features and occupancy prediction.

[0065] The above map uncertainty feature modeling and representation includes the following steps:

[0066] D1: Position uncertainty modeling: Online vector map estimation usually uses regression methods to predict the positions of polygon vertices, expressed as position vectors v=v1,v2) in a two-dimensional BEV coordinate system. A regression head containing position and scale parameters is used to convert this prediction result into a probability model. Assume that each map element consists of multiple vertices, and the vertex set is represented as in is the position of the i-th vertex. For each vertex position, the Laplace distribution is used to describe its uncertainty. The distribution formula is:

[0067]

[0068] in, is the expected value of the position, is the scale parameter, which indicates the uncertainty of the jth coordinate of the i-th vertex;

[0069] D2: Constructing a joint probability density function: A map element M containing multiple vertices comprehensively describes the uncertainty of the entire map element by combining the uncertainties of all vertices, so that the overall uncertainty of the online map is better characterized. The formula for the joint probability density function is:

[0070]

[0071] D3: Probabilistic modeling of the output of the category classifier. The classification head will output the confidence score of the category to which each vertex belongs. The confidence score is the probability distribution of the classifier. For each vertex v (i) , let its category label be c (i) ,The uncertainty of the category is quantified by the distribution of classification confidence, so that the entire map element has uncertainty measures in terms of both location and category;

[0072] D4: Design a loss function to replace the traditional mean square error regression loss with negative log likelihood NLL loss to capture the uncertainty of the position, so that the real distribution of the data can be better reflected in the model prediction process. The loss function formula is:

[0073]

[0074] This loss function encourages the model to optimize both the expected and uncertainty parameters of the location.

[0075] The above-mentioned fusion of multi-camera 3D occupancy prediction includes the following steps:

[0076] E1: Multi-camera feature extraction and representation: Given a set of multi-camera image inputs {I 1 ,I 2 ,…,I N}, a backbone network is used to extract the features of each camera, and the feature extraction of each camera produces a multi-scale feature representation, denoted as in represents the features of the jjjth scale from the iiith camera. These features capture different levels of information of the scene, from low-level edge information to high-level semantic information;

[0077] E2: Adopting the Transformer-based spatial cross attention mechanism, using the spatial relationship between multiple cameras to weight the fusion features, given the features from different cameras These features are aggregated through cross-viewpoint cross-attention to propagate information between different cameras, thereby obtaining a three-dimensional convolutional volume feature V, ensuring that the generated 3D occupied volume can maximize the fusion of the spatial information of each camera. The formula of the three-dimensional convolutional volume feature V is:

[0078] V=G(I 1 ,I 2 ,…,I N );

[0079] Where G is a neural network module that uses multi-head cross attention to map the 2D features of multiple cameras into 3D occupancy probabilities;

[0080] E3: 2D to 3D spatial mapping: When implementing 3D occupancy prediction, 2D features are fused through a 2D to 3D mapping operation. Using the deformation attention mechanism, 3D voxel reference points are projected onto the 2D view and features around these 2D points are sampled for fusion. For each 3D query point q p , project it into the corresponding 2D view, and the output features It is the result of weighted summation based on the cross-view attention mechanism:

[0081]

[0082] in, represents the set of viewing angles from which the 3D reference point can be observed, Q p is the 3D volume query of the pth element, and DeformAttn represents the weighted feature calculated by deformation attention;

[0083] E4: Multi-scale feature fusion and prediction: Make full use of multi-level feature information, adopt 2D-3D U-Net architecture, extract and fuse 3D convolution volume features at multiple scales through 2D to 3D spatial mapping operations, and extract and fuse 3D convolution volume features at multiple scales into given multi-scale 2D features. After multiple 2D-3D spatial attention layers, multi-scale 3D convolution features are obtained Then the 3D convolutional features of the j-1th layer are upsampled and fused with the features of the jth layer to obtain Y j , the formula is:

[0084] Y j =F j +Deconv(Y j-1 );

[0085] By combining high-level semantic features and low-level detail features, occupancy prediction results with different resolutions are generated.

[0086] E5: Determine the loss function and supervision mechanism: The model can be effectively supervised at different scales, and a supervision signal is introduced at each scale. Specifically, for the occupancy prediction results of each scale, the present invention uses cross entropy loss and affinity loss of scene categories as supervision signals, and the loss function is:

[0087]

[0088] in, To attenuate the loss weight, it is more important to ensure high-resolution prediction. Through this multi-scale supervision mechanism, the network can learn high-level semantic features and detail features at the same time, thereby improving the accuracy of 3D scene reconstruction;

[0089] Through the above-mentioned multi-camera feature extraction, multi-scale fusion, spatial cross-attention and supervision mechanism, the present invention achieves more accurate environmental modeling in multi-camera 3D occupancy prediction and improves the autonomous driving perception capability in complex scenarios.

[0090] The joint modeling of map uncertainty features and occupancy predictions includes the following steps:

[0091] F1: Encode map features: Encode the uncertainty of each map element. Each map element M is composed of several vertices. Let the position of the i-th vertex be v (i) , the uncertainty is given by the location parameter μ (i) and scale parameter b (i) Indicates that the category probability c is introduced (i) , used to describe the category information of map elements, combine the above information, encode the characteristics of each vertex, and the encoding result is expressed as:

[0092]

[0093] in is the uncertainty characteristic of the map;

[0094] F2: Joint modeling and fusion: Combine the occupancy grid feature V obtained by multi-camera fusion, take the uncertainty information of the map as a priori, and jointly perform occupancy prediction. Let the fused occupancy grid feature be V = [v1, v2, …, v M ], each feature v j Corresponding to a voxel in a three-dimensional space, the uncertainty characteristics of the map Fusion into the occupancy prediction of each voxel to enhance the prediction of occupancy state;

[0095] For the occupancy state prediction of each voxel, the following formula is used for joint modeling:

[0096]

[0097] Among them, f() is a joint modeling function, which uses linear weighting or nonlinear functions to fuse the occupancy features from multiple cameras and the uncertainty features of the map;

[0098] The present invention can use a linear fusion function to weight the sum of map priors and occupancy features, and then limit the output range through an activation function:

[0099]

[0100] Here, σ(·) is the activation function, and w1 and w2 are learnable parameters used to adjust the influence of different information sources. In this way, the model can combine the uncertainty information of the map in the occupancy state prediction, thereby outputting more conservative estimates in areas with higher uncertainty, improving the system's adaptability to dynamic environments.

[0101] F3: Design loss functions to enable the model to effectively learn joint modeling: Establish multi-task loss functions Including cross entropy loss for occupancy states Regularization loss fused with uncertainty features The formula is:

[0102]

[0103] in is the cross entropy loss of the occupied state, is the regularization loss of uncertainty fusion, λ is the trade-off coefficient, Used to measure the error between the predicted voxel occupancy state and the true annotation, To ensure that the model does not overfit to sensor noise or defective areas when considering map uncertainty information, λ is used to balance the contribution of occupancy prediction error and uncertainty regularization to the overall loss.

[0104] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A look-around occupancy grid method combined with online vector map feature priors, characterized in that: The steps include: S1: Detailed modeling of online map uncertainty features, that is, jointly modeling the uncertainty of map element positions and categories to provide more robust occupancy predictions, using a Laplace distribution-based approach to quantify the uncertainty in vectorized map features, and completing the modeling and representation of map uncertainty features; S2: Using multi-camera input data, by fusing the multi-scale features of each camera, to achieve more accurate 3D occupancy prediction, and complete the fusion of multi-camera 3D occupancy prediction; S3: The uncertainty features in the map are used as prior information and jointly modeled with the occupancy prediction task to complete the joint modeling of map uncertainty features and occupancy prediction.

2. The method of looking around and occupying a grid in combination with online vector map feature priors according to claim 1, characterized in that: The map uncertainty feature modeling and representation includes the following steps: D1: Position uncertainty modeling: Online vector map estimation usually uses regression methods to predict the positions of polygon vertices, which are expressed as position vectors v = (v1, v2) in the two-dimensional BEV coordinate system. A regression head containing position and scale parameters is used to convert this prediction result into a probability model. Assume that each map element consists of multiple vertices, and the vertex set is represented as in is the position of the i-th vertex. For each vertex position, the Laplace distribution is used to describe its uncertainty. The distribution formula is: in, is the expected value of the position, is the scale parameter, which indicates the uncertainty of the jth coordinate of the i-th vertex; D2: Constructing a joint probability density function: A map element M containing multiple vertices is used to comprehensively describe the uncertainty of the entire map element by combining the uncertainties of all vertices, so that the overall uncertainty of the online map is better characterized; D3: Probabilistic modeling of the output of the category classifier. The classification head outputs the confidence score of the category to which each vertex belongs. The confidence score is the probability distribution of the classifier. For each vertex v (i) , let its category label be c (i) ,The uncertainty of the category is quantified by the distribution of classification confidence, so that the entire map element has uncertainty measures in terms of both location and category; D4: Design a loss function to replace the traditional mean square error regression loss with negative log likelihood (NLL) loss to capture the uncertainty of the position.

3. The method of combining online vector map feature prior with look-around occupancy grid according to claim 2, characterized in that: The formula of its joint probability density function is:

4. The method of combining online vector map feature prior with look-around occupancy grid according to claim 3, characterized in that: The loss function formula is: This loss function encourages the model to optimize both the expected and uncertainty parameters of the location.

5. The method of looking around and occupying a grid in combination with online vector map feature priors according to claim 1, characterized in that: The fusion of multi-camera 3D occupancy prediction includes the following steps: E1: Multi-camera feature extraction and representation: Given a set of multi-camera image inputs {I 1 , I 2 , ..., I N }, a backbone network is used to extract the features of each camera, and the feature extraction of each camera produces a multi-scale feature representation, denoted as in represents the features of the jjjth scale from the iiith camera; E2: Adopting the Transformer-based spatial cross attention mechanism, using the spatial relationship between multiple cameras to weight the fusion features, given the features from different cameras By cross-attention, the features Aggregate to obtain a three-dimensional convolutional volume feature V; E3: 2D to 3D spatial mapping: When implementing 3D occupancy prediction, 2D features are fused through a 2D to 3D mapping operation. Using the deformation attention mechanism, 3D voxel reference points are projected onto the 2D view and features around these 2D points are sampled for fusion. For each 3D query point q p , project it into the corresponding 2D view, and the output features It is the result of weighted summation based on the cross-view attention mechanism: in, represents the set of viewing angles from which the 3D reference point can be observed, Q p is the 3D volume query of the pth element, and DeformAttn represents the weighted feature calculated by deformation attention; E4: Multi-scale feature fusion and prediction: We make full use of multi-level feature information and adopt the 2D-3DU-Net architecture to extract and fuse 3D convolutional volume features at multiple scales through 2D to 3D spatial mapping operations; E5: Determine the loss function and supervision mechanism: The model can be effectively supervised at different scales, and a supervision signal is introduced at each scale. Specifically, for the occupancy prediction results of each scale, the present invention uses cross entropy loss and affinity loss of scene categories as supervision signals, and the loss function is: in, is the attenuated loss weight.

6. The method of combining online vector map feature prior with look-around occupancy grid according to claim 5, characterized in that: The formula for the three-dimensional convolution volume feature V is: V=G(I 1 ,I 2 ,...,I N ); Where G is a neural network module that uses multi-head cross attention to map the 2D features of multiple cameras into 3D occupancy probabilities.

7. The method of combining online vector map feature prior with look-around occupancy grid according to claim 6, characterized in that: Extract and fuse 3D convolutional volume features at multiple scales into 2D features at given multiple scales After multiple 2D-3D spatial attention layers, multi-scale 3D convolution features are obtained Then the 3D convolutional features of the j-1th layer are upsampled and fused with the features of the jth layer to obtain Y j , the formula is: Y j =F j +Decony(Y j-1 ); By combining high-level semantic features and low-level detail features, occupancy prediction results with different resolutions are generated.

8. The method of combining online vector map feature prior with look-around occupancy grid according to claim 1, characterized in that: The joint modeling of map uncertainty features and occupancy predictions includes the following steps: F1: Encode map features: Encode the uncertainty of each map element. Each map element M is composed of several vertices. Let the position of the i-th vertex be v (i) , the uncertainty is given by the location parameter μ (i) and scale parameter b (i) Indicates that the category probability c is introduced (i) , used to describe the category information of map elements, combine the above information, encode the characteristics of each vertex, and the encoding result is expressed as: in is the uncertainty characteristic of the map; F2: Joint modeling and fusion: Combine the occupancy grid feature V obtained by multi-camera fusion, take the uncertainty information of the map as a priori, and jointly perform occupancy prediction. Let the fused occupancy grid feature be V = [v1, v2, ..., v M ], each feature v j Corresponding to a voxel in a three-dimensional space, the uncertainty characteristics of the map Fusion into the occupancy prediction of each voxel to enhance the prediction of occupancy state; For the occupancy state prediction of each voxel, the following formula is used for joint modeling: Among them, f() is a joint modeling function, which uses linear weighting or nonlinear functions to fuse the occupancy features from multiple cameras and the uncertainty features of the map; F3: Design loss functions to enable the model to effectively learn joint modeling: Establish multi-task loss functions Including cross entropy loss for occupancy states Regularization loss fused with uncertainty features The formula is: in is the cross entropy loss of the occupied state, is the regularization loss of uncertainty fusion, and λ is the trade-off coefficient.

9. The method of combining online vector map feature prior with look-around occupancy grid according to claim 8, characterized in that: In step F2, a linear fusion function is also used to weight the map prior and occupancy features, and then the output range is limited by the activation function: Among them, σ(·) is the activation function, w1 and w2 are learnable parameters used to adjust the influence from different information sources.

10. The method of combining online vector map feature prior with look-around occupancy grid according to claim 8, characterized in that: Said Used to measure the error between the predicted voxel occupancy state and the true annotation, To ensure that the model does not overfit to sensor noise or defective areas when considering map uncertainty information, λ is used to balance the contribution of occupancy prediction error and uncertainty regularization to the overall loss.

Images