Methods and systems for adapting pre-trained BEV networks and pre-trained BEV networks

By using a pre-trained base model and a differentiable neural renderer, the multi-view BEV network is automatically adapted, solving the problem of 3D information errors caused by changes in sensor configuration. This achieves efficient, low-cost adaptation and accurate 3D perception, making it suitable for autonomous driving and robotics.

CN123071791APending Publication Date: 2026-11-10ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610615952.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2025-05-07
Filing Date
2026-05-07
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing multi-view BEV networks exhibit incorrect orientation and scaling of predicted 3D information when sensor configuration changes, affecting path planning and collision avoidance performance. Furthermore, existing adaptation methods are costly and lack generalization ability.

Method used

By leveraging a pre-trained base model and a differentiable neural renderer, multi-view BEV networks are automatically adapted to new sensor configurations through depth estimation and error calculation, and end-to-end optimization is performed using unlabeled real data.

Benefits of technology

It achieves efficient and low-cost multi-view BEV network adaptation under new sensor configurations, improves the accuracy and robustness of 3D perception, and is applicable to autonomous driving and robotics technologies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071791A_ABST
    Figure CN123071791A_ABST
Patent Text Reader

Abstract

This invention relates to a method for adapting a pre-trained BEV network to a new sensor configuration, particularly a new camera device configuration, wherein the method includes: - acquiring multiple environmental datasets, particularly multiple images, from the new sensor configuration; - generating a target depth map for each of the multiple environmental datasets, with depth estimation performed using a pre-trained base model; - generating a predicted 3D occupancy display based on the multiple images, with the pre-trained BEV network in use; - generating a predicted depth map for each of the multiple environmental datasets based on the predicted 3D occupancy display, particularly with the use of a differentiable neural renderer; - determining an error function and / or an error magnitude based on a comparison between the target depth map and the predicted depth map, respectively; and - adapting the pre-trained BEV network based on the determined error magnitude and / or error function. Furthermore, the invention also relates to a corresponding system and a pre-trained BEV network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to methods and systems for adapting pre-trained multi-view bird's-eye view (BEV) networks to new sensor configurations. The invention also relates to a pre-trained multi-view bird's-eye view (BEV) network for environmental recognition. Background Technology

[0002] Multi-view BEV networks are crucial for various 3D perception tasks, but they are particularly sensitive to biases in camera configurations between training and use. Applying a network trained for a specific camera configuration to other configurations results in predicted 3D information, including grid occupancy and object positions, exhibiting incorrect orientation in height and depth, as well as incorrect scaling. These inaccuracies can significantly impair the performance and reliability of subsequent tasks such as path planning and collision avoidance.

[0003] Adapting multi-view bird's-eye view (BEV) networks to new sensor configurations, especially new camera configurations, typically requires costly and time-consuming detection and annotation of data for the target configuration. Existing solutions attempt to overcome this limitation through different strategies. One strategy involves training with diverse datasets that include different camera configurations. However, the sheer number of potential camera locations and orientations makes it practically impossible to build sufficiently diverse datasets to cover all real-world scenes. Therefore, these approaches often fail to generalize effectively to significantly different sensor geometries. Another approach uses synthetic data generation, where a pre-trained generative model, such as StableDiffusion, is used along with geometric transformations to synthesize images from the new camera's perspective. While promising, this approach faces several challenges. The computational cost of generating large synthetic datasets can be prohibitive. Furthermore, synthetic images often differ stylistically from real data in terms of lighting, object representation, and sensor noise, which impairs the effectiveness of training with such data. This "simulation-to-real" gap frequently leads to poor generalization to real-world scenes. Summary of the Invention

[0004] Therefore, the objective of this invention can be considered as providing a more robust, efficient, and cost-effective method for adapting a pre-trained multi-view bird's-eye view (BEV) network to a new sensor configuration.

[0005] According to a first aspect of the invention, a method is proposed for adapting a pre-trained multi-view bird's-eye view (BEV) network to new sensor configurations, particularly new camera device configurations. The method includes the following steps: - Acquire multiple environmental datasets, particularly multiple images, from the new sensor configuration; - Generate a target depth map for each of the multiple environment datasets, using a pre-trained base model (FM) for depth estimation; - When using the pre-trained BEV network, a predicted 3D occupancy display is generated based on the multiple images; - Based on the predicted 3D occupancy display, generate a predicted depth map for each of the multiple environment datasets, especially when using a differentiable neural renderer. - Determine the error function and / or error amount based on the comparison between the target depth maps and the predicted depth maps that are respectively associated with each other; - Adapt the pre-trained multi-view bird's-eye view (BEV) network based on the determined error function and / or error amount.

[0006] This invention is based on recent developments in the field of foundational models (FM) for depth estimation. A key feature of this foundational model is its significant accuracy in predicting depth maps based on a single image. This capability is crucial because it provides easily accessible supervisory signals for the adaptation process. The reliance on explicit benchmark truth parallax data, which is typically unavailable or prohibitively expensive to obtain, is eliminated.

[0007] In the context of this invention, a "base model" (FM) refers to a large, pre-trained deep learning model, particularly one trained for depth estimation based on a single image. Such models are typically trained on massive datasets, such as datasets of images and corresponding depth information, enabling them to learn general features and patterns in depth perception. A core feature of the base model in this context is its ability to generalize well to new, unseen data or images and to accurately predict depth maps without further fine-tuning or adaptation to the target dataset. This invention leverages this generalization ability by using the base model's depth predictions as a supervisory signal to adapt a multi-view BEV network to new camera configurations, thereby eliminating the need for labeled depth data based on the target structure. Although the base model itself is not specifically trained for the target task (adaptation of the BEV network), its pre-trained depth estimation capabilities provide valuable information that controls the adaptation process. Examples of basic models suitable for this purpose are Metric3DV2 (see Hu et al., “Metric3Dv2: A Versatile Monocular Geometric Foundation Model for Zero-shot Metric Depth and Surface Normal Estimation”, https: / / arxiv.org / abs / 2404.15506) and DepthAnything (see Yang et al., “Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data”, https: / / arxiv.org / abs / 2401.10891), both of which are trained on large datasets and exhibit excellent performance in monocular depth estimation.

[0008] In the context of this invention, a "multi-view bird's-eye-view (BEV) network" refers to a specialized artificial neural network developed for 3D perception based on multiple datasets of environmental data, such as images from camera devices. This network can, for example, receive images as input from multiple camera devices mounted at different viewpoints around the vehicle (or other platform), and can transform these perspective views into a unified bird's-eye-view display. This relates to orthographic projection, i.e., a perspective-distortion-free map-like top-down view, which presents the scene as if viewed from above. This display is typically created in a fixed coordinate system centered on the vehicle.

[0009] The key lies in the network learning process: fusing information from multiple sensor or camera perspectives to construct a coherent and comprehensive 3D understanding of the scene within the BEV space. This 3D understanding can take the form of a 3D occupancy grid, object recognition within the BEV space, or other 3D features relevant to subsequent tasks, such as autonomous driving. The network architecture typically includes: encoders for processing various sets of environmental data (e.g., camera images); transformations for projecting features into the BEV space; and a fusion mechanism for combining multi-view information. The network is typically trained simultaneously using data from multiple camera devices to learn these transformations and fusion operations.

[0010] In the context of this invention, a "differentiable neural renderer" is a key component that enables the computation of the gradient of the loss relative to the parameters of the BEV network. Specifically, the differentiable neural renderer receives a 3D occupancy display predicted by the BEV network as input and generates a synthetic depth map based on this, which simulates the scene from the perspective of a single camera device. In this respect, "differentiable" means that the operation of the renderer is constrained in such a way that the gradient of the error (the difference between the rendered depth map and the depth map predicted by the base model) relative to the input data (3D occupancy display) can be computed. This gradient computation is particularly crucial for the backpropagation method, which is used to adjust the parameters of the BEV network during training. Thus, the differentiable neural renderer enables the BEV network to be trained in such a way that it produces a 3D occupancy display consistent with the depth information of the base model, without requiring labeled data based on a new camera device configuration. The differentiable neural renderer acts as a bridge between the 3D output of the BEV network and the 2D depth map, and therefore enables end-to-end optimization of the entire system.

[0011] The sensor can be, first and foremost, a variety of general-purpose environmental sensors, such as radar sensors, lidar sensors, ultrasonic sensors, or imaging sensors (especially camera devices). Here, the sensor configuration can include multiple sensors of the same type, or it can include different sensor types.

[0012] For example, a sensor configuration may include multiple camera devices. A new sensor configuration may, for example, include more and / or different sensors compared to the original sensor configuration. The sensors in the new sensor configuration may be oriented or configured differently compared to the sensors in the original sensor configuration.

[0013] This method is specifically designed to adapt a pre-trained multi-view bird's-eye view (BEV) network to new camera device configurations, wherein the method includes the following steps: - Provides a pre-trained base model for depth estimation and a differentiable neural renderer; - Acquire multiple images from the new camera device configuration, especially unlabeled images; - Use the pre-trained base model (FM) to generate a target depth map for each image; - Use the pre-trained multi-view bird's-eye view (BEV) network to generate a predicted 3D occupancy display based on the multiple images; - Use the differentiable neural renderer to generate a predicted depth map for each image based on the predicted 3D occupancy display; - Based on the difference between the target depth map and the predicted depth map, determine the error function and / or the error amount; and - Adapt the pre-trained multi-view bird's-eye view (BEV) network based on the determined error function and / or error amount.

[0014] In a preferred embodiment of the invention, the generation of the predicted depth map includes projecting the predicted 3D occupancy display onto a corresponding image plane of the new camera device configuration, using the differentiable neural renderer. This provides the advantage of "accurately taking into account the specific camera device geometry." This results in improved accuracy of 3D perception and enables precise localization and classification of objects in a vehicle environment.

[0015] In another preferred embodiment, if the pre-trained base model provides disparity information with invariant scaling, then an error function and / or an error magnitude with invariant scaling is selected as the error function. Conversely, if the pre-trained base model provides absolute metric disparity information, then an absolute error function (such as mean squared error, MSE) is selected. Adapting the pre-trained multi-view bird's-eye view (BEV) network involves fine-tuning the pre-trained BEV network based on the error magnitude thus determined using backpropagation. This constrained selection of the error function based on the characteristics of the base model (invariant scaling or metric) ensures robust adaptation and stable convergence of the training process. Fine-tuning the network using backpropagation enables automated optimization of all relevant network parameters and results in optimal adaptation to new sensor configurations.

[0016] Particularly preferably, environmental datasets, especially images from the camera device, are provided without labeling. The method according to the invention can, in principle, be performed with labeled or unlabeled datasets. However, a particular advantage of this method is demonstrated by the fact that labeling is not required. Therefore, existing unlabeled data, which can be provided, for example, by the calibration of the new sensor configuration, can be advantageously utilized. This significantly reduces the overhead for data detection and preprocessing and accelerates the adaptation process, leading to cost savings and faster development.

[0017] In another preferred embodiment, the 3D occupancy display is generated as the main output by the pre-trained multi-view bird's-eye view (BEV) network.

[0018] In another alternative preferred embodiment, the multi-view bird's-eye view (BEV) network generates a 3D occupancy display as an auxiliary output, and additionally performs at least one additional perception task on top of this. The perception task may in particular include 3D object recognition and / or lane recognition and / or semantic segmentation. This allows the multi-view bird's-eye view (BEV) network to be integrated into more complex systems and enables more complete and detailed environmental perception.

[0019] In another preferred embodiment, adapting the pre-trained multi-view bird's-eye view (BEV) network includes adjusting the 3D transformation that maps image features to BEV space. This allows for precise compensation of geometric deviations between old and new sensor configurations. This improves the system's robustness to changes in sensor layout and results in reliable 3D perception even when sensor arrangement changes.

[0020] In another preferred embodiment, adapting the pre-trained multi-view bird's-eye view (BEV) network includes adjusting the previously learned feature representations of the pre-trained BEV network. Adjusting the learned feature representations allows for finer-tuning based on new sensor configurations. This leads to improved generalization to new, unknown scenarios and increased system reliability in real-world applications.

[0021] In another preferred embodiment, the plurality of environmental datasets or the plurality of images come from different sensors in the new sensor configuration, particularly different camera devices in the new camera device configuration.

[0022] In another preferred embodiment, the method is performed iteratively, wherein the steps of "generating a predicted depth map," "determining a loss value," and "fitting the pre-trained multi-view bird's-eye view (BEV) network" are repeated in each iteration. This allows for progressive optimization of the BEV network and improves the convergence speed of the learning process, leading to more efficient fitting and better results.

[0023] According to a second aspect of the invention, a system is provided for adapting a pre-trained multi-view bird's-eye view (BEV) network to a new camera device configuration. The system is particularly configured for implementing the method according to the first aspect. The system includes: The memory stores: - A pre-trained base model for depth estimation, configured to generate a target depth map based on an image; and - A differentiable neural renderer configured to create predicted depth maps based on 3D occupancy displays.

[0024] The system also includes a processor configured to: - Acquire multiple images from the new camera device configuration; - This causes the pre-trained base model to generate a target depth map for each of the plurality of images; - Obtain a predicted 3D occupancy display generated by the pre-trained multi-view bird's-eye view (BEV) network based on the multiple images; - This prompts the differentiable neural renderer to create a predicted depth map for each of the plurality of images based on the predicted 3D occupancy display; - Based on the difference between the target depth map and the predicted depth map, determine the error amount and / or error function; and - Based on the determined error amount and / or error function, output the training signal for the pre-trained multi-view bird's-eye view (BEV) network.

[0025] Preferably, the system is configured such that the processor is further configured to iteratively perform the steps of “generating a predicted depth map”, “determining a loss value”, and “adapting the pre-trained BEV network”.

[0026] According to a third aspect, a pre-trained multi-view bird's-eye view (BEV) network is proposed for environmental recognition, particularly for environmental recognition of the environment of at least partially automated motor vehicles or the environment of autonomous robots. The network includes at least one module for generating a 3D occupancy display based on received environmental image data acquired using a determined sensor configuration. The BEV network is configured to output a description of the environment, particularly a list of 3D objects and / or an occupancy grid, based on the 3D occupancy display generated by the module. According to the invention, the BEV network is configured to adapt to new sensor configurations, according to the method described in the first aspect, particularly using the system described in the second aspect.

[0027] Particularly preferably, the pre-trained multi-view bird's-eye view (BEV) network is integrated into autonomous or partially automated vehicles or autonomous robots. Therefore, the pre-trained BEV network enables the flexible and efficient adaptation of perception systems (such as those commonly used in modern vehicles or robots) to new sensor configurations and tasks, thereby enhancing the autonomy and performance of such systems and enabling a wide range of applications.

[0028] Therefore, this invention provides a novel approach to adapting multi-view BEV networks to new sensor or camera configurations without requiring newly labeled data for this purpose. Unlike previous approaches that typically only adapt portions of the network, comprehensive end-to-end fine-tuning is performed here. A pre-trained base model (FM) for depth estimation is used, which serves as supervision during the adaptation process. Thus, not only are the network branches used for depth estimation optimized, but also the key 3D transformations that convert image features to BEV space, and all implicitly learned feature representations. This holistic optimization ensures seamless adaptation of the entire perception pipeline to new sensor geometries. In a particularly preferred configuration of this invention, neural rendering is used. This generates depth maps based on 3D occupancy predicted by the network. These depth maps form a bridge between the 3D occupancy output and the parallax supervision provided by the FM.

[0029] The methods and / or systems according to the invention can be used in vehicles. The vehicle can be configured, for example, as a motor vehicle and / or a passenger vehicle and / or at least a partially automated / autonomous vehicle. The vehicle may have, for example, vehicle facilities for providing autonomous driving functions and / or driver assistance systems. These vehicle facilities can be implemented for, at least partially automatically, controlling and / or accelerating and / or braking and / or steering the vehicle. A multi-view BEV perception system can be used for at least one driving function. The sensor arrangement may be part of the vehicle or arranged on the vehicle.

[0030] The method and system according to the present invention can significantly reduce the cost and time overhead of adapting perception models to new sensor arrangements. Expensive reference sensor systems and manual annotation are no longer required. Compared to existing methods, especially those based on diverse datasets or synthetic data, the present invention provides higher accuracy and robustness. This is achieved by utilizing real sensor data from the target configuration, thereby directly correcting for deviations caused by variations in the geometry of the camera device.

[0031] This invention has broad applications in every multi-view BEV perception system that must be adapted to new sensor configurations. While the method can be directly applied to 3D occupancy prediction, it can be seamlessly integrated into other tasks such as 3D object recognition, lane recognition, and semantic segmentation. In these cases, 3D occupancy prediction can be integrated as an auxiliary output, where the primary task implicitly benefits from an improved 3D transformation learned during adaptation. This is achieved by utilizing the calibrated transformation across tasks, resulting in more accurate and geometrically consistent outputs. This invention is particularly effective in applications such as driver assistance, autonomous driving, robotics, and general 3D scene understanding. Attached Figure Description

[0032] The embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0033] Figure 1 A flowchart illustrating a method according to an embodiment of the present invention is shown.

[0034] Figure 2 A block diagram illustrating an exemplary network architecture of a pre-trained multi-view bird's-eye view (BEV) network, and a system according to another embodiment of the present invention for adapting a pre-trained multi-view bird's-eye view (BEV) network to a new camera device configuration.

[0035] In the following description of embodiments of the invention, the same elements are designated by the same reference numerals, and repeated descriptions of these elements will be omitted where necessary. These figures are merely schematic illustrations of the subject matter of the invention. Detailed Implementation

[0036] Figure 1 Flowchart 500 illustrates the steps of a method according to the invention for adapting a pre-trained multi-view bird's-eye view (BEV) network to a new camera device configuration. In step 510, the process begins with data detection, where images of the environment, such as a vehicle environment, are detected using the new camera device configuration. Subsequently, in step 520, inference of the base model is performed, wherein a target depth map is generated for each image using the pre-trained base model. Then, or in parallel, in step 530, the pre-trained BEV network generates a predicted 3D occupancy display based on the input image. In step 540, a predicted depth map is generated for each image based on the 3D occupancy display using a differentiable neural renderer. In the next step 550, an error function and / or error magnitude are calculated by comparing the target depth map with the predicted depth map. Finally, in step 560, the pre-trained BEV network is adapted based on the calculated error function and / or error magnitude. This adaptation can be performed iteratively.

[0037] Figure 2 A block diagram illustrating an exemplary network architecture of a pre-trained multi-view bird's-eye view (BEV) network 200 is shown, illustrating how the invention can be integrated and implemented using a system 300 constructed as a sensor generalization module. The core architecture 200 of the BEV network receives input images 110 from a new camera device configuration and processes these input images via an encoder 210 and an image-to-BEV transformer 220 to extract BEV features 230. Based on this, a 3D occupancy prediction is generated in module 240, which forms the basis for subsequent tasks 250 such as 3D object recognition, lane recognition, or semantic segmentation. During the retraining phase, the actual adaptation according to the new sensor configuration is performed in the system 300, constructed as a sensor generalization module, for adapting the pre-trained multi-view bird's-eye view (BEV) network 200 to the new camera device configuration. The system 300 utilizes a base model 310 to generate a target depth map based on the input images 110. Differentiable neural renderer 320 generates a corresponding predicted depth map based on the 3D occupancy prediction 240 of the BEV network. Based on an error function and / or error magnitude, the target depth map is compared with the predicted depth map in training loop 330, and a training signal 140 is generated based on this comparison to adjust the parameters of the BEV network 200.

[0038] Therefore, system 300 embodies the core functionality of the invention and enables adaptation of BEV network 200 without labeled data. After this adaptation, system 300 can be removed, and the adapted BEV network 200 can independently provide the desired result 130.

Claims

1. A method for adapting a pre-trained multi-view bird's-eye view (BEV) network (200) to new sensor configurations, particularly new camera device configurations, wherein, The method includes the following steps: - Acquire multiple environmental datasets, in particular multiple images (110), from the new sensor configuration; - Generate a target depth map for each of the multiple environment datasets, using a pre-trained base model (310) for depth estimation; - When using the pre-trained multi-view bird's-eye view (BEV) network (200), a predicted 3D occupancy display is generated based on the multiple images (110); - Based on the predicted 3D occupancy display, generate a predicted depth map for each of the multiple environment datasets, especially when using a differentiable neural renderer (320). - Based on a comparison between the target depth maps and the predicted depth maps, respectively, determine the error function and / or error magnitude; and - The pre-trained multi-view bird's-eye view (BEV) network (200) is adapted based on the determined error amount and / or error function.

2. The method according to claim 1, wherein, The generation of the predicted depth map includes projecting the predicted 3D occupancy display onto the corresponding image plane of the new camera device configuration, using the differentiable neural renderer.

3. The method according to claim 1 or 2, wherein, - If the pre-trained base model (FM) provides depth information with scale invariant, then the error function with scale invariant is selected as the error function; or, if the pre-trained base model (FM) provides depth information with absolute metric, then the absolute error function and / or error amount, such as the mean square error (MSE) error function, is selected. as well as - Adapting the pre-trained multi-view bird's-eye view (BEV) network includes: fine-tuning the pre-trained multi-view bird's-eye view (BEV) network based on a determined error function when using backpropagation.

4. The method according to any one of claims 1 to 3, wherein, Provide an environmental dataset from the new sensor configuration without labeling.

5. The method according to any one of the preceding claims, wherein, The pre-trained multi-view bird's-eye view (BEV) network generates a 3D occupancy display as the main output.

6. The method according to any one of claims 1 to 4, wherein, The pre-trained multi-view bird's-eye view (BEV) network generates a 3D occupancy display as an auxiliary output and additionally performs at least one other perception task, particularly 3D object recognition and / or lane recognition and / or semantic segmentation.

7. The method according to any one of the preceding claims, wherein, The adaptation of the pre-trained multi-view bird's-eye view (BEV) network includes adjusting the 3D transformation that maps image features to the BEV space.

8. The method according to any one of the preceding claims, wherein, Adapting the pre-trained multi-view bird's-eye view (BEV) network includes adjusting the feature representations learned by the pre-trained multi-view bird's-eye view (BEV) network.

9. The method according to any one of the preceding claims, wherein, The multiple environmental datasets, and in particular the multiple images, come from different sensors, and in particular different camera devices, configured with the new sensor configuration, and in particular the new camera device configuration.

10. The method according to any one of the preceding claims, wherein, The method is performed iteratively, wherein in each iteration the steps of "generating a predicted depth map", "determining the error function and / or error amount" and "adapting the pre-trained multi-view bird's-eye view (BEV) network" are repeated.

11. A system for adapting a pre-trained multi-view bird's-eye view (BEV) network (200) to a new camera device configuration, particularly according to the method of any one of claims 1 to 10, wherein, The system (300) includes: The memory stores: - A pre-trained base model (310) for depth estimation, configured to generate a target depth map based on an image; and - A differentiable neural renderer (320) configured to create a predicted depth map based on a 3D occupancy display; and A processor configured to: - Acquire multiple images (110) from the new camera device configuration; - This causes the pre-trained base model (310) to generate a target depth map for each of the plurality of images (110); - Obtain a predicted 3D occupancy display (240) generated by the pre-trained multi-view bird's-eye view (BEV) network (200) based on the multiple images (110); - This prompts the differentiable neural renderer (320) to create a predicted depth map for each of the plurality of images (110) based on the predicted 3D occupancy display; - Based on the difference between the target depth map and the predicted depth map, determine the error amount and / or error function; and - Based on the determined error amount and / or error function, output the training signal (140) for the pre-trained multi-view bird's-eye view (BEV) network (200).

12. The system according to claim 11, wherein, The processor is also configured to iteratively perform the steps of "generating a predicted depth map", "determining the amount of error and / or the error function", and "adapting the pre-trained multi-view bird's-eye view (BEV) network".

13. A pre-trained multi-view bird's-eye view (BEV) network (200) for environment recognition, particularly for environment recognition of at least partially automated motor vehicles or autonomous robots, said pre-trained multi-view bird's-eye view (BEV) network comprising: A module (240) for generating a 3D occupancy display based on received environmental image data (110), said image data being acquired with the aid of a determined sensor configuration; The multi-view bird's-eye view (BEV) network (200) is configured to output a description of the environment, particularly a list of 3D objects and / or an occupancy grid, based on a 3D occupancy display generated by the module (240). The multi-view bird's-eye view (BEV) network (200) can be adapted to a new sensor configuration according to the method of any one of claims 1 to 10, especially by means of the system according to claim 11 or 12.

14. The pre-trained multi-view bird's-eye view (BEV) network (200) for environment recognition according to claim 13, wherein, The pre-trained multi-view bird's-eye view (BEV) network is integrated into autonomous vehicles or autonomous robots.