Vehicle aerial view three-dimensional detection method based on adaptive long-distance environmental perception enhancement

Abstract

The invention relates to a vehicle aerial view three-dimensional detection method based on adaptive long-distance environmental perception enhancement, multi-view image data of a vehicle is used as input of a detection model to obtain a detection result, and a processing process of the detection model comprises the following steps: performing feature extraction on the multi-view image data by using a backbone network to obtain initial image features; based on the preliminary image features and branch-level dynamic learning weights, fusing multi-scale local information and global information, and combining a region attention mechanism and a pixel attention mechanism of adaptive weights to obtain multi-view image features; based on the multi-view image features, mapping the multi-view image features to an aerial view space through view conversion, and obtaining an aerial view feature map by combining aerial view query, spatial cross attention and time self-attention; and based on the aerial view feature map, predicting a 3D target frame and a speed, and realizing vehicle aerial view three-dimensional detection. According to the invention, adaptive learning and regulation of the scene and the target level are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, and in particular to a vehicle bird's eye view detection method based on adaptive long-distance environment perception enhancement Figure Three . BACKGROUND

[0002] With the rapid development of automatic driving technology, as one of the core modules, environment perception puts forward higher requirements for the accurate detection and positioning of targets in the three-dimensional space around the vehicle. Traditional two-dimensional target detection methods are difficult to effectively obtain the depth information and spatial structure in the scene, and are difficult to meet the needs of automatic driving for three-dimensional understanding. Therefore, three-dimensional target detection technology has gradually become a research hotspot. Existing three-dimensional detection methods mainly include two categories based on laser radar and camera images. The laser radar method has high spatial accuracy, but the sensor cost is high and is not easy to deploy on a large scale. The camera method has the advantages of low cost and rich information, and has unique advantages in long-distance perception and semantic understanding. Therefore, three-dimensional target detection methods based on pure vision have received widespread attention in recent years. Bird's-Eye-View (BEV) as a unified and spatially aligned representation has been widely used in three-dimensional detection tasks in multi-camera systems. By converting multi-view image features to BEV space, the detection performance and system efficiency are improved to a certain extent.

[0003] However, the existing BEV detection method still faces two key challenges: first, the features of long-distance targets are prone to degradation, resulting in a significant decline in detection performance; second, the multi-scale feature fusion method is relatively static, and it is difficult to dynamically adjust the information weight according to the scene changes, affecting the adaptability in complex environments. In addition, some methods rely heavily on depth information, which makes errors easily accumulate in the spatial mapping process, thereby affecting the overall detection accuracy. SUMMARY

[0004] The purpose of the present application is to overcome the defects of the prior art and provide a vehicle bird's eye view detection method based on adaptive long-distance environment perception enhancement Figure Three to solve or partially solve the problems that existing methods cannot adaptively learn and regulate different scenes and target scales, and have weak adaptability to complex driving scenes.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] In one aspect of the present application, a vehicle bird's eye view detection method based on adaptive long-distance environment perception enhancement Figure Three is provided, which takes multi-view image data of a vehicle as input of a detection model to obtain a detection result, and the processing process of the detection model includes the following steps:

[0007] extracting features from the multi-view image data by using a backbone network to obtain preliminary image features;

[0008] fusing multi-scale local information and global information based on the preliminary image features and dynamic learning weights of branch levels, combining an adaptive weight region attention mechanism and a pixel attention mechanism to obtain multi-view image features;

[0009] based on the multi-view image features, mapping to a bird's eye view space through view conversion, combining bird's eye view queries, spatial cross-attention, and temporal self-attention to obtain a bird's eye view feature map;

[0010] based on the bird's eye view feature map, predicting a 3D target frame and a speed to realize vehicle bird's eye view Figure Three detection.

[0011] As a preferred technical solution, the loss weight of the target at different depth distances is dynamically adjusted through adaptive depth perception balance, a loss function is calculated based on the loss weight, and the training of the detection model is realized.

[0012] As a preferred technical solution, the loss function is:

[0013] Cost=reg_cost*ω(d)+cls_cost*ω(d)

[0014]

[0015] wherein Cost is a loss function, reg_cost is a regression loss of a bounding box, cls_cost is a classification loss of a bounding box, ω(d) is a depth dynamic adjustment weight, distance i is a depth distance of the i-th target depth, α is an initial distance weight of the interval where the target depth distance is located, θ is a distance adjustment factor, and Max() represents calculating the maximum value for all i.

[0016] As a preferred technical solution, the fusion of multi-scale local information and global information based on the preliminary image features and dynamic learning weights of branch levels, combining an adaptive weight region attention mechanism and a pixel attention mechanism to obtain multi-view image features includes the following steps:

[0017] extracting local features F Context and global features F Spatial from the preliminary image features, and obtaining preliminary fusion features F fuse by superimposition;

[0018] The preliminary fusion feature is divided into multiple non-overlapping regions of different scales, and average pooling, channel transformation and up-sampling are performed to form a multi-channel region attention feature

[0019] The preliminary fusion feature is compressed and expanded to obtain a pixel attention feature

[0020] Based on the adaptive weight β, the region attention feature and the pixel attention feature are fused to obtain a fusion feature F R+P ;

[0021] Based on the fusion feature F R+P , the weight α of each position in the feature map is calculated;

[0022] Based on the weight α, the local feature F Context and the global feature F Spatial are fused to obtain a multi-view image feature.

[0023] As a preferred technical solution, the multiple different scales include 1×1, 2×2 and 4×4.

[0024] As a preferred technical solution, the multi-view image data covers different view regions around the vehicle.

[0025] As a preferred technical solution, the spatial cross-attention is:

[0026]

[0027] Wherein, F t is the feature of the t-th view in the multi-view image feature {F1,..., F n}, n is the total number of views, Q p is the three-dimensional column query result obtained by querying the bird's eye view, SCA(Q p , F t ) is the spatial cross-attention result, V hit is N ref 3D reference points sampled from the three-dimensional column, and the hit view obtained by projecting to the 2D image view, P(p,i,j) represents the position of the j-th reference point on view i, is the feature map of view i, and DeformAttn() represents the deformable attention mechanism.

[0028] As a preferred technical solution, the time self-attention is:

[0029]

[0030] Wherein, Q pfor a query point Q at a spatial coordinate p in the bird's eye view space, B' t-1 p t-1 t-1

[0031] In another aspect of the present application, an electronic device is provided, comprising: one or more processors and a memory having stored therein one or more programs, the one or more programs including instructions for performing the aforementioned vehicle bird's eye view detection method based on adaptive long-range environment perception enhancement. Figure Three

[0032] In another aspect of the present application, a computer-readable storage medium is provided, comprising one or more programs for execution by one or more processors of an electronic device, the one or more programs including instructions for performing the aforementioned vehicle bird's eye view detection method based on adaptive long-range environment perception enhancement. Figure Three

[0033] Compared with the prior art, the present application has at least one of the following beneficial effects:

[0034] (1) Realize the adaptive learning and regulation of scene and target level: the present application provides a multi-scale adaptive attention fusion scheme, which fuses multi-scale local information and global information based on preliminary image features and dynamic learning weights of branch level, combines regional attention mechanism and pixel attention mechanism of adaptive weight to obtain multi-view image features, can finely model and selectively fuse feature information of different scales, improve the accuracy and robustness of target detection, and modify the MSA part to add dynamic learning weights of different branches, aiming to realize selective learning between different scale branches, so that the model can dynamically adjust the fusion proportion of different scale branches according to the loss feedback in the training process, realize adaptive learning and regulation of different scenes and target scales.

[0035] (2) Strong adaptability to complex driving scenes: the present application provides a distance-based deep weight dynamic adjustment mechanism, which dynamically adjusts the loss weight of different depth distance targets through adaptive depth perception balance, calculates the loss function based on the loss weight, effectively solves the problems of insufficient optimization of long-distance targets and uneven distribution of loss weights in traditional methods, not only improves the detection accuracy of long-distance targets by the mechanism, but also considers the training stability of close-range targets, thereby enhancing the overall detection performance of the model and its adaptability to complex driving scenes.​​​​​​ BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 Schematic diagram of vehicle bird's eye view in the embodiment Figure Three Flowchart of the method for detecting the vehicle;

[0037] Figure 2 Schematic diagram of multi-scale attention part in the embodiment

[0038] Figure 3 Schematic diagram of multi-scale adaptive attention feature fusion in the embodiment

[0039] Figure 4 Schematic diagram of the electronic device in the embodiment. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor should fall within the scope of protection of the present application.

[0041] Embodiment 1

[0042] In view of the problems of the prior art described above, the present embodiment provides a vehicle bird's eye view enhanced based on adaptive long-distance environment perception Figure Three The method for detecting the vehicle, please refer to Figure 1 , the method comprises the following steps:

[0043] Step S1, collect multi-view camera image data from an autonomous driving sensor system, the camera is arranged at multiple positions around the vehicle, for covering different view angle areas around the vehicle, to obtain complete environment perception image information.

[0044] Step S2, provide the multi-camera image to a backbone network (such as ResNet-50, ResNet-101), at time t, and obtain features of different camera views (B,C,H,W), wherein B represents the number of images (or feature maps) input to the model at a time, C represents the number of channels, H represents the height of the image or feature map, W represents the width of the image or feature map, the superscript i represents the i-th view, N view represents the total number of camera views.

[0045] Step S3, continue multi-scale self-adaptive attention feature extraction (MSAAF) for multi-view image features, mainly divided into two parts: multi-scale attention (MSA) and multi-scale attention fusion (MSAF), in addition, this step adds dynamic learning weights (β1, β2, β3, β4) of different branches in the MSA part, as shown in Figure 2 , aiming to realize selective learning between different scale branches, so that the model can adapt to different scenes and targets, and strengthen the adaptive ability of the model. The core idea is to let the network learn feature weights according to the loss, so that the model can selectively fuse information from different scales (in the formula, context represents local features, and spatial represents global features). Local features can be extracted through shallow high-resolution branches, using methods such as 3x3 convolution, small receptive field or local self-attention mechanism, to capture detailed and small target information. Global features come from deep low-resolution branches, obtained through large receptive field, global average pooling or global attention mechanism, focusing on overall semantics and structural relationships.

[0046] F Fuse = F context+ F Spatial

[0047] In the formula, F Fuse , that is, F(CxHxW) in Figure 2 .

[0048] Step S301, multi-scale attention (MSA).

[0049] As shown in Figure 2 , the main purpose of MSA is to learn the weight α as the basis for fusing different levels of branches. MSA is mainly divided into two parts: region attention (Region Attention) and pixel attention (Pixel Attention). In order to explain MSA in more detail, the number of channels, the width and the height of the feature map are denoted as C, W and H respectively.

[0050] (1) Region attention.

[0051] Region Attention aims to enhance the model's ability to capture multi-scale spatial dependencies while reducing computational complexity. Unlike traditional pixel-level or channel-level attention, Region Attention divides the feature map into blocks of different sizes and calculates attention weights independently within each region to highlight key spatial information. To adapt to feature extraction of targets of different scales, the input feature map is divided into multiple non-overlapping regions of different scales, such as 1x1, 2x2, and 4x4. For the divided region (N x N), first, average pooling is performed to obtain a compact regional feature representation. Similar to the SE attention mechanism, the pooled regional features are then compressed and expanded in the channel to optimize their information expression ability. To ensure that the region attention is compatible with the pixel-level attention, the feature map after channel transformation compression (GONV) is restored to the original resolution for weighted fusion with the input features. In the formula, unpool represents the up-sampling operation, and avgpool represents the average pooling operation.

[0052]

[0053] (2) Pixel Attention.

[0054] Pixel Attention is used to measure the importance of each pixel. This module does not need to perform pooling or deformation operations and can run directly on the original spatial resolution. This allows the model to focus on key areas while maintaining the integrity of spatial information. As shown in the following formula, the original feature F fuse is directly compressed and expanded in the channel to obtain F

[0055]

[0056] Finally, each feature is first concatenated along the channel dimension and then normalized using Softmax to calculate the adaptive weight β i . The calculated region attention weight is multiplied element-wise with the original feature map, as shown in the following formula, to enhance the focus on key regions while suppressing background or irrelevant information, where F R+P represents the sum of the features of region attention and pixel attention.

[0057]

[0058] In the above formula, F R+P is the feature after channel compression and expansion in F Figure 2 . fuse

[0059] Step S302, Multi-Scale Attention Fusion (MSAF). ​

[0060] By superimposing pixel attention and region attention, the weights of different positions in the feature map are obtained. As shown in Figure 3 , we finally fuse the two branches by element-wise multiplication. The mathematical formula of region attention is shown in the following equation, where sigmoid represents the activation function, and add represents feature splicing, Figure 2 represents the global feature extraction branch in Figure 2 represents the multi-scale features of each branch in

[0061]

[0062] F Out =F Context ×α+F Spatial ×(1-α)

[0063] Step S4, input the multi-view image features after feature extraction into the view conversion module based on the Transformer structure, unify the features from different cameras into a unified bird's eye view (BEV) space, and output the BEV feature map representing the environment structure.

[0064] The main architecture of the Transformer view conversion module has 6 layers of traditional transformer structure encoder. It mainly consists of three parts: BEV query (BEV Queries), spatial cross-attention (SCA) and temporal self-attention (TSA).

[0065] (1) BEV Queries.

[0066] It includes a set of BEV queries Q ∈ R C×H×W , which is used to sample and fuse information in the BEV space. BEV queries are similar to target queries in DETR (DEtection TRansformer), but they are not used for target detection, but for BEV representation learning.

[0067] (2) Spatial Cross-Attention (SCA)

[0068] In order to aggregate information from multi-view image features {F1,...,F n}, the main use is spatial cross-attention mechanism. SCA aggregates information by calculating the attention score of query Q and image feature F n . The attention mechanism can be represented by the following formula:​​

[0069]

[0070] First, a three-dimensional columnar query Q p is made, and N ref 3D reference points are sampled from the column, and these 3D reference points are projected into 2D image views, only part of the views can receive the projected points, called "hit views" (V hit ), and features are sampled at the 2D positions of the hit views, P(p,i,j) represents the position of the jth reference point on view i, is the feature map of view i, and DeformAttn represents the deformable attention mechanism.

[0071] (3) Temporal Self-Attention (TSA).

[0072] In order to consider the timing information, a time cross attention (TSA) mechanism is introduced, which allows the BEV query of the current frame to interact with the BEV query of the historical frame. The time attention mechanism can be represented as:

[0073]

[0074] For a query point Q p located at the BEV space coordinate p, TSA needs to integrate the BEV query Q of the current timestamp and the aligned historical BEV feature B' t-1 to extract features. First, the historical BEV feature B t-1 is converted to B' t-1 by self-motion alignment, so that it is aligned with the current BEV query in the same coordinate system. Then, TSA adopts a deformable attention (Deformable Attention) mechanism to perform attention calculation between the current BEV query Q and the historical BEV feature B' t-1 to capture the temporal correlation. Among them, DeformAttn(Q p , p, V) selectively aggregates features from V by predicting dynamic offsets, ensuring that the model can focus on the change information of moving objects at different time steps.

[0075] Step S5, the bird's eye view features are input into the detection head module, and the detection head directly predicts the 3D target frame and the speed, replacing the traditional 2D frame regression. Through this design, the model can end-to-end predict the 3D target frame and the speed, without relying on non-maximum suppression (NMS) for post-processing, so as to realize more concise and efficient extraction and three-dimensional target detection related feature information.

[0076] Step S6, in the model training process, an adaptive depth perception balancing mechanism is used to dynamically adjust the loss weight of different depth distance targets, aiming to enhance the detection ability of the model for long-distance targets, thereby improving the overall three-dimensional target detection performance. The specific steps are as follows:

[0077] First, the depth range is divided into multiple intervals (for example, near (0-20), medium (20-35), and far (35-50)), and different loss weights [2.0 3.0 5.0] are applied to different intervals. An initial distance weight a is assigned to the target, and then an exponential function is used (theta is a distance adjustment factor that can control the speed of exponential growth, distance i is the depth distance of the ith target).

[0078]

[0079] Finally, the loss function is:

[0080] Cost=reg_cost*ω(d)+cls_cost*ω(d);

[0081] In the formula, reg_cost is the regression loss of the bounding box, cls_cost is the classification loss of the bounding box, omega(d) is the depth dynamic adjustment weight, and Cost is the overall loss.

[0082] Step S7, based on the detection head output and the optimization result of the loss function, a three-dimensional detection box result is generated for target positioning and recognition.

[0083] Embodiment 2

[0084] The embodiment provides an electronic device, comprising one or more processors and a memory, the memory storing one or more programs, the one or more programs comprising instructions for performing a vehicle bird's eye view Figure Three detection method based on adaptive long-distance environment perception enhancement as described in embodiment 1.

[0085] As Figure 2 described, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and of course, other hardware required for business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs to implement the above Figure 1 described method. Of course, in addition to the software implementation, the present application does not exclude other implementation manners, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or a logic device.

[0086] Memory can include non-persistent memory in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read only memory (ROM) or flash memory, etc. Memory is an example of computer-readable media.

[0087] Embodiment 3

[0088] This embodiment provides a computer-readable storage medium including one or more programs for execution by one or more processors of an electronic device, the one or more programs including instructions for performing a vehicle bird's-eye view Figure Three detection method as described in Embodiment 1.

[0089] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology for storing information. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage device, or any other non-transmission medium that can be used to store information accessible by a computing device. According to the definition herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carriers.

[0090] The present application not only effectively deals with the problem of feature weakening of long-distance targets in three-dimensional target detection, but also introduces a multi-scale dynamic feature fusion mechanism to improve the perception ability of the model in different scales and complex scenes. By constructing a unified representation in the bird's-eye view space and combining a depth adaptive loss function to optimize the detection weight of near and far targets, the target recognition ability of the model in long-distance environments is significantly enhanced. The three-dimensional detection model trained based on this method performs excellently in multi-view pure visual perception systems and has a wide application prospect, including, for example:

[0091] (1) Autonomous driving environment perception, this method can be used to build a high-precision three-dimensional perception system, improve the detection accuracy of vehicles for long-distance pedestrians, vehicles and obstacles in complex road environments, and enhance the path planning and decision-making ability;

[0092] (2) Intelligent traffic monitoring system, suitable for urban road monitoring equipment, through enhancing the long-distance target detection capability, realizing efficient perception and management of large-scale traffic targets, and improving the urban traffic operation efficiency and safety;

[0093] (3) Robot navigation and perception, which can be integrated into a mobile robot platform to realize fast and accurate detection and obstacle avoidance of key targets in multi-scale scenes, and support autonomous navigation tasks in complex indoor and outdoor environments.

[0094] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement, characterized in that, The multi-view image data of the vehicle is taken as the input of a detection model to obtain a detection result, and the processing procedure of the detection model comprises the following steps: A backbone network is used to extract features from the multi-view image data to obtain preliminary image features; Based on the preliminary image features and dynamic learning weights of branch levels, multi-scale local information and global information are fused, and adaptive weight region attention mechanism and pixel attention mechanism are combined to obtain multi-view image features; Based on the multi-view image features, a bird's eye view space is mapped through perspective conversion, bird's eye view queries, spatial cross attention, and time self-attention are combined to obtain a bird's eye view feature map; Based on the bird's eye view feature map, 3D target boxes and speeds are predicted to realize vehicle bird's eye view three-dimensional detection.

2. The vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement according to claim 1, characterized in that, The loss weights of targets at different depth distances are dynamically adjusted through adaptive depth perception balance, a loss function is calculated based on the loss weights to realize training of the detection model.

3. The vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement according to claim 2, characterized in that, The loss function is: Cost=reg_cost*ω(d)+cls_cost*ω(d) wherein Cost is a loss function, reg_cost is a regression loss of the bounding box, cls_cost is a classification loss of the bounding box, ω(d) is a depth dynamic adjustment weight, distance i is a depth distance of the i-th target depth, a is an initial distance weight of an interval in which the target depth distance is located, θ is a distance adjustment factor, Max() represents calculating all i Then the maximum value is taken.

4. The vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement according to claim 1, characterized in that, Based on the preliminary image features and dynamic learning weights of branch levels, multi-scale local information and global information are fused, and adaptive weight region attention mechanism and pixel attention mechanism are combined to obtain multi-view image features, which comprises the following steps: Based on the preliminary image features, local features F are extracted Context and global features F Spatial , and the preliminary fusion features F fuse are obtained by superposition. The preliminary fusion features are divided into multiple non-overlapping regions of different scales and are subjected to average pooling, channel transformation and up-sampling to form multi-channel regional attention features the preliminary fusion features are obtained by compression and expansion to obtain pixel attention features Fusing the region attention feature and the pixel attention feature based on an adaptive weight β to obtain a fused feature F R+P ; based on the fusion feature F R+P , calculate the weight α of each position in the feature map; Based on the weight α, the local feature F Context and the global feature F Spatial are fused to obtain a multi-view image feature.

5. The vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement according to claim 4, characterized in that, The multiple different scales comprise 1×1, 2×2, and 4×4.

6. The vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement according to claim 1, characterized in that, The multi-view image data covers different perspective regions around the vehicle.

7. The vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement according to claim 1, characterized in that, The spatial cross attention is: Among them, F t For multi-view image features {F1,...,F n The features of the t-th viewpoint in}, where n is the total number of viewpoints, Q p The 3D column query results obtained from the bird's-eye view query are SCA(Q). p ,F t For example, V represents the spatial cross-attention result. hit To sample N from a 3D cylinder ref A 3D reference point is generated and then projected onto a 2D image view to obtain the hit view. P(p,i,j) represents the position of the j-th reference point on view i, and F... t i It is the feature map of view i, and DeformAttn() represents the deformable attention mechanism.

8. The vehicle bird's eye view 3D detection method based on adaptive long-range environment perception enhancement according to claim 1, characterized in that, The time self-attention is: where Q p is the query point at spatial coordinate p in the bird's eye view, Q, B t-1 are the current timestamp's bird's eye view query and aligned historical BEV features, respectively, TSA(Q p , B t-1 ) is the temporal self-attention result, DeformAttn() denotes the deformable attention mechanism that selectively aggregates features from V by predicting dynamic offsets, V denotes the set containing elements Q, B t-1 .

9. An electronic device, comprising: It comprises: One or more processors and memories, the memories store one or more programs, the one or more programs comprise instructions for executing the vehicle bird's eye view three-dimensional detection method based on adaptive long-distance environmental perception enhancement according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It comprises one or more programs for one or more processors of an electronic device to execute, the one or more programs comprise instructions for executing the vehicle bird's eye view three-dimensional detection method based on adaptive long-distance environmental perception enhancement according to any one of claims 1-8.

Images