Lane line detection method and system combining standard definition map and satellite map

Abstract

The invention provides a lane line detection method and system combining a standard definition map and a satellite map, and the method comprises the steps: carrying out the coding of an SD map image through employing a polyline sequence representation method, and obtaining the features of the SD map; inputting the multi-view image of the vehicle-mounted camera into the image backbone network for feature extraction to obtain image features and image BEV features, and sequentially performing attention mechanism processing and feature fusion to obtain BEV features; acquiring satellite map image features by using a U-net network, and performing alignment and fusion on the satellite map image features and the BEV features to obtain final fusion features; and decoding the final fusion feature to generate a lane topology prediction result. According to the method, the positions of the multi-segment line points are embedded and coded by using the sine with different frequencies, so that the sensitivity of the model to the position change is improved, and the accuracy of lane line reasoning is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and computer vision analysis, and in particular to a lane line detection method and system that combines standard definition maps with satellite maps. Background Art

[0002] As the main means of transportation in road traffic, cars are an indispensable part of the transportation field. However, the high number of cars has aggravated traffic congestion, seriously affected the safety and operating efficiency of the transportation system, and also led to the deterioration of the traffic environment. In recent years, the number of people killed or disabled in road traffic accidents has continued to rise worldwide, especially in complex road conditions such as bad weather and night driving, the incidence of traffic accidents is much higher than under normal conditions.

[0003] In order to effectively prevent traffic accidents and improve the safety performance of vehicles, the need to develop and promote intelligent safety systems is becoming more urgent, such as electronic stability control systems (ESC), anti-lock braking systems (ABS) and advanced driver assistance systems (ADAS). Among them, ADAS can not only reduce the risk of accidents by continuously assisting drivers in completing driving tasks, but also realize autonomous driving functions. In the development of autonomous driving technology and ADAS systems, lane line detection is a crucial task. Its goal is to accurately identify and locate lane lines on the road to help vehicles stay in lane and make safe driving decisions.

[0004] Although the current lane detection technology has shown high detection accuracy in good lighting and simple road conditions, problems still remain in complex scenarios. For example, the lane line may be unclear due to being blocked or worn by other vehicles, or in harsh driving conditions such as rainy days, nights, and heavy fog, the on-board multi-view camera is affected by the environment, resulting in blurred images, which in turn reduces the accuracy of lane detection. Summary of the invention

[0005] In view of the above situation, the main purpose of the present invention is to propose a lane line detection method and system that combines standard definition maps and satellite maps to solve the above technical problems.

[0006] The present invention proposes a lane line detection method combining standard definition map and satellite map fusion, the method comprising the following steps: Step 1: Acquire images based on the open lane line dataset to obtain multi-view images of the vehicle camera, SD map images, and satellite map images; Step 2: construct an extraction network based on the Transformer encoder, and use the polyline sequence representation method to encode the SD map image to obtain the SD map features; Step 3: Based on the extraction network, the multi-view image of the vehicle camera is input into the image backbone network for feature extraction to obtain image features and image BEV features; The image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing to obtain the processed image BEV features; and the processed image BEV features are fused with the SD map features using the multi-head cross attention mechanism to obtain the BEV features; Step 4: Use the U-net network to obtain satellite map image features, align and fuse the satellite map image features and BEV features to obtain the final fusion features; Step 5: Decode the final fused features to generate lane topology prediction results.

[0007] The present invention also proposes a lane line detection system combining standard definition map and satellite map fusion, the system comprising: Content input module for: Image acquisition is performed based on the open lane line dataset to obtain multi-view images from vehicle-mounted cameras, SD map images, and satellite map images; SD map feature fusion module, used for: An extraction network is constructed based on the Transformer encoder, and the SD map image is encoded using a polyline sequence representation method to obtain SD map features. Based on the extraction network, the multi-view images of the vehicle camera are input into the image backbone network for feature extraction to obtain image features and image BEV features; The image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing to obtain the processed image BEV features; and the processed image BEV features are fused with the SD map features using the multi-head cross attention mechanism to obtain the BEV features; Satellite map feature fusion module, used for: Use the U-net network to obtain satellite map image features, align and fuse satellite map image features and BEV features to obtain the final fusion features Lane topology prediction output module, used for: The final fused features are decoded to generate lane topology prediction results.

[0008] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention improves the sensitivity of the model to position changes and enhances the accuracy of lane line reasoning by using sinusoidal embedding with different frequencies to encode the positions of multi-segment line points; 2. The present invention uses one-hot vectors to represent the main lane types, which ensures the normalization of input values ​​and solves the problem that some road segments may belong to multiple types; 3. The present invention uses step-by-step fusion and mask attention to refine features in the satellite map feature fusion module, ensuring high-quality feature alignment and obtaining accurate final fusion features; 4. The present invention can provide efficient and real-time lane prediction in complex environments and harsh conditions, improve road perception and topology understanding capabilities, and is particularly suitable for lane line detection in complex scenarios.

[0009] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flowchart of the steps of the lane line detection method that combines standard definition maps and satellite maps proposed in the present invention.

[0011] Figure 2 This is the overall architecture diagram of the lane line detection method that combines standard definition maps and satellite maps proposed in the present invention.

[0012] Figure 3 This is a structural diagram of the SD map feature extraction module of the lane line detection method that combines standard definition maps and satellite maps proposed in the present invention.

[0013] Figure 4 This is a structural diagram of the satellite map feature step-by-step fusion module of the lane line detection method that combines standard definition maps and satellite maps proposed in the present invention.

[0014] Figure 5 This is a technical test comparison chart of the lane line detection method that combines standard definition maps and satellite maps proposed in the present invention.

[0015] Figure 6 This is a structural diagram of the lane line detection system that combines standard definition maps and satellite maps proposed in the present invention. DETAILED DESCRIPTION

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout are the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0017] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to provide some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0018] See also Figure 1 The embodiment of the present invention proposes a lane line detection method combining standard definition map and satellite map fusion, the method comprising the following steps: Step 1: Acquire images based on the open lane line dataset to obtain multi-view images of the vehicle-mounted camera, SD map images, and satellite map images.

[0019] Furthermore, in this embodiment, an open lane line dataset is used as a training image input to the network, and all images are Fixed size.

[0020] Step 2: Build an extraction network based on the Transformer encoder, use the polyline sequence representation method to encode the SD map image, and obtain the SD map features.

[0021] See also Figure 2 and Figure 3 In step 2, an extraction network is constructed based on the Transformer encoder, and the SD map image is encoded using a polyline sequence representation method to obtain SD map features, which specifically includes the following steps: Segment sampling is performed on the SD map image to obtain a sampled target, and the sampled target is encoded by sinusoidal embedding to obtain an encoding result; Perform one-hot encoding on the lane type to obtain a label representing the road type; The encoding result is concatenated with the one-hot vector representing the road type label to obtain the SD map feature.

[0022] In the step of performing line segment sampling on the SD map image to obtain the sampled target and encoding the sampled target through sine embedding, the corresponding relationship between the process is as follows: ; in, Indicates that after embedding, represents the sine function, represents the cosine function, Indicates the coordinate position, represents the frequency index, represents the influence of temperature scale, represents the temperature scale, represents the scale factor, represents the embedding dimension, Represents division.

[0023] In the step of one-hot encoding the lane type, the corresponding relationship is as follows: ; in, represents the one-hot encoding function, represents the Kroneth function, Indicates the lane category that currently needs to be coded. represents the total number of lane categories, represents the first category in the lane category set, Indicates the lane category set categories.

[0024] Step 3: Based on the extraction network, the multi-view image of the vehicle camera is input into the image backbone network for feature extraction to obtain image features and image BEV features; The image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing to obtain the processed image BEV features; and the processed image BEV features are fused with the SD map features using the multi-head cross attention mechanism to obtain the BEV features.

[0025] In step 3, the image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing. The corresponding relationship of the process is as follows: ; in, It means that it has been processed by the spatial cross attention mechanism. represents the image BEV feature query sequence, a key sequence representing image features, A sequence of values ​​representing image features, represents the scaling numerator; In the step of fusing the processed image BEV features with the SD map features using the multi-head cross attention mechanism, the corresponding relationship between the process is as follows: ; in, represents the multi-head spatial cross self-attention mechanism function, Indicates the corresponding The mapping matrix of the image BEV feature query matrix is: Indicates the corresponding The mapping matrix of the SD map feature key matrix, Indicates the corresponding The mapping matrix of the SD map eigenvalue matrix, represents the weight matrix, Indicates a connection operation. Indicates branches, represents the query sequence of BEV features processed by the spatial cross-attention mechanism, A key sequence representing SD map features, A sequence of values ​​representing SD map features.

[0026] It should be noted that the BEV feature represents a bird's-eye view feature.

[0027] Step 4: Use the U-net network to obtain satellite map image features, align and fuse the satellite map image features and BEV features to obtain the final fused features.

[0028] See also Figure 4 In step 4, the U-net network is used to obtain satellite map image features, and the satellite map image features and BEV features are aligned and fused to obtain the final fusion features, which specifically includes the following steps: Based on the extraction network, a step-by-step fusion module is constructed, which includes a feature-level fusion module and a BEV-level fusion module. For satellite map images, the Unet network is used to perform dilated convolution processing to obtain satellite map image features; The feature-level fusion module is used to mix the BEV features and satellite map image features through learnable parameters, and the distance mask and segmentation mask are used to adjust the position and segment the information respectively to obtain the refined BEV features. Combine the distance mask and the segmentation mask to get the final mask; The outputs of a preset number of masked cross attention modules are connected in sequence to obtain a cascaded masked cross attention module, and the refined BEV features are processed by the cascaded masked cross attention module and the attention mechanism is combined with the final mask to obtain the refined BEV features processed by the mechanism; The refined BEV features processed by the mechanism are calculated through a feed-forward network to obtain refined BEV features; The BEV-level fusion module is used to perform a bilinear interpolation kernel transformation operation on the satellite map image features and the refined BEV features to obtain the aligned satellite map features; The aligned satellite map features are fused with the refined BEV features to obtain the final fused features.

[0029] In the step of using the Unet network to perform dilated convolution processing on satellite map images, the corresponding relationship between the process is as follows: ; in, Represents the pixel value of the output satellite map feature map, Represents the index position of the output feature map in the vertical direction, Represents the index position in the horizontal direction of the output feature map, represents the half width of the convolution kernel, represents the sampling points after expansion in the input satellite map feature map, represents the expansion rate, Represents the index of the convolution kernel in the vertical direction, Represents the index of the convolution kernel in the horizontal direction, Represents the weight of the convolution kernel.

[0030] In the step of combining the distance mask and the segmentation mask to obtain the final mask, the corresponding relationship between the process is as follows: ; in, represents the final mask, represents the segmentation mask, Represents the distance mask; In the step of processing the refined BEV features through the cascaded mask cross attention module and combining the final mask with the attention mechanism, the corresponding relationship in the process is as follows: ; in, represents the masked cross attention function, represents the normalization operation, represents the query sequence of BVE features processed by the multi-head spatial cross self-attention mechanism, The transpose of the key sequence representing the satellite map, A sequence of values ​​representing a satellite map, Indicates transpose.

[0031] In the step of calculating the refined BEV features processed by the mechanism through the feedforward network to obtain the refined BEV features, the corresponding relationship in the process is as follows: ; in, represents the refined BEV features, It means that after the feed-forward network processing, represents the masked criss-cross attention function.

[0032] In the step of using the BEV-level fusion module to perform a bilinear interpolation kernel transformation operation on the satellite map image features and the refined BEV features to obtain the aligned satellite map features, the corresponding relationship of the process is as follows: ; in, Represents the aligned satellite map features, Indicates the horizontal coordinate of the satellite map feature map, Indicates the vertical coordinate of the satellite map feature map, Indicates the height of the satellite map feature map, Indicates the width of the satellite map feature map. Represents the horizontal coordinates of adjacent pixels used to calculate bilinear interpolation, Represents the ordinates of adjacent pixels used to calculate bilinear interpolation, Represents the characteristics of satellite map images, represents the first two-dimensional transformation offset learned for the position, Represents the second two-dimensional transformation offset learned for the position, Indicates taking the maximum value.

[0033] Furthermore, the distance mask is used for position adjustment, and the distance mask has the following relationship: ; in, represents the Euclidean distance function, represents the distance threshold, Represents the position coordinates in the hybrid BEV space that combines the SD map image features with the on-board camera image features.

[0034] The segmentation mask uses a fully convolutional layer to generate the semantic segmentation information of the satellite map, and then uses linear embedding and expands its dimension to obtain the segmentation mask.

[0035] It should be noted that in Figure 4 In the paper, the flow field is a set of two-dimensional offset vectors generated by a convolutional network, and bilinear interpolation is used to generate aligned features. The purpose is to align satellite maps with BEV features, solve the positioning error problem, enhance the quality of multimodal fusion, and make the fused features have better geometric consistency, thus improving the model's ability in distance prediction and occluded area reasoning.

[0036] Step 5: Decode the final fused features to generate lane topology prediction results.

[0037] See also Figure 5In order to prove that the present invention has excellent performance compared with the prior art, the present invention is tested and compared with the prior art model in this embodiment. The test adopts an open lane line data set. The test results are as follows: Figure 6 As shown in the figure, the evaluation indicators are: average accuracy of directional lane centerlines, average accuracy of topological relationships between lane centerlines, average accuracy of traffic elements, average accuracy of topological relationships between lane centerlines and traffic elements, and overall lane centerline evaluation metric. Figure 5 The larger the evaluation index value in , the higher the accuracy of topological reasoning and the better the model performance.

[0038] See also Figure 6 The present invention also provides a lane line detection system combining standard definition map and satellite map fusion, the system comprising: Content input module for: Image acquisition is performed based on the open lane line dataset to obtain multi-view images from vehicle-mounted cameras, SD map images, and satellite map images; SD map feature fusion module, used for: An extraction network is constructed based on the Transformer encoder, and the SD map image is encoded using a polyline sequence representation method to obtain SD map features. Based on the extraction network, the multi-view images of the vehicle camera are input into the image backbone network for feature extraction to obtain image features and image BEV features; The image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing to obtain the processed image BEV features; and the processed image BEV features are fused with the SD map features using the multi-head cross attention mechanism to obtain the BEV features; Satellite map feature fusion module, used for: The U-net network is used to obtain satellite map image features, and the satellite map image features and BEV features are aligned and fused to obtain the final fusion features; Lane topology prediction output module, used for: The final fused features are decoded to generate lane topology prediction results.

[0039] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0040] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0041] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A lane line detection method combining standard definition map and satellite map fusion, characterized in that: The method comprises the following steps: Step 1: Acquire images based on the open lane line dataset to obtain multi-view images of the vehicle camera, SD map images, and satellite map images; Step 2: construct an extraction network based on the Transformer encoder, and use the polyline sequence representation method to encode the SD map image to obtain the SD map features; Step 3: Based on the extraction network, the multi-view image of the vehicle camera is input into the image backbone network for feature extraction to obtain image features and image BEV features; The image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing to obtain the processed image BEV features; and the processed image BEV features are fused with the SD map features using the multi-head cross attention mechanism to obtain the BEV features; Step 4: Use the U-net network to obtain satellite map image features, align and fuse the satellite map image features and BEV features to obtain the final fusion features; Step 5: Decode the final fused features to generate lane topology prediction results.

2. The lane line detection method combining standard definition map and satellite map fusion according to claim 1, characterized in that: In step 4, the satellite map image features are obtained by using the U-net network, and the satellite map image features and the BEV features are aligned and fused to obtain the final fused features, which specifically includes the following steps: Based on the extraction network, a step-by-step fusion module is constructed, which includes a feature-level fusion module and a BEV-level fusion module. For satellite map images, the Unet network is used to perform dilated convolution processing to obtain satellite map image features; The feature-level fusion module is used to mix the BEV features and satellite map image features through learnable parameters, and the distance mask and segmentation mask are used to adjust the position and segment the information respectively to obtain the refined BEV features. Combine the distance mask and the segmentation mask to get the final mask; The outputs of a preset number of masked cross attention modules are connected in sequence to obtain a cascaded masked cross attention module, and the refined BEV features are processed by the cascaded masked cross attention module and the attention mechanism is combined with the final mask to obtain the refined BEV features processed by the mechanism; The refined BEV features processed by the mechanism are calculated through a feed-forward network to obtain refined BEV features; The BEV-level fusion module is used to perform a bilinear interpolation kernel transformation operation on the satellite map image features and the refined BEV features to obtain the aligned satellite map features; The aligned satellite map features are fused with the refined BEV features to obtain the final fused features.

3. The lane line detection method combining standard definition map and satellite map fusion according to claim 2 is characterized in that: In the step of using the Unet network to perform dilated convolution processing on satellite map images, the corresponding relationship between the process is as follows: ； in, Represents the pixel value of the output satellite map feature map, Represents the index position of the output feature map in the vertical direction, Represents the index position in the horizontal direction of the output feature map, represents the half width of the convolution kernel, represents the sampling points after expansion in the input satellite map feature map, represents the expansion rate, Represents the index of the convolution kernel in the vertical direction, Represents the index of the convolution kernel in the horizontal direction, Represents the weight of the convolution kernel.

4. The lane line detection method combining standard definition map and satellite map fusion according to claim 3 is characterized in that: In the step of combining the distance mask and the segmentation mask to obtain the final mask, the corresponding relationship between the process is as follows: ； in, represents the final mask, represents the segmentation mask, Represents the distance mask; In the step of processing the refined BEV features through the cascaded mask cross attention module and combining the final mask with the attention mechanism, the corresponding relationship in the process is as follows: ； in, represents the masked cross attention function, represents the normalization operation, represents the query sequence of BVE features processed by the multi-head spatial cross self-attention mechanism, The transpose of the key sequence representing the satellite map, A sequence of values ​​representing a satellite map, represents transpose; In the step of calculating the refined BEV features processed by the mechanism through the feedforward network to obtain the refined BEV features, the corresponding relationship in the process is as follows: ； in, represents the refined BEV features, It means that after the feed-forward network processing, represents the masked criss-cross attention function.

5. The lane line detection method combining standard definition map and satellite map fusion according to claim 4 is characterized in that: In the step of using the BEV-level fusion module to perform a bilinear interpolation kernel transformation operation on the satellite map image features and the refined BEV features to obtain the aligned satellite map features, the corresponding relationship of the process is as follows: ； in, Represents the aligned satellite map features, Indicates the horizontal coordinate of the satellite map feature map, Indicates the vertical coordinate of the satellite map feature map, Indicates the height of the satellite map feature map, Indicates the width of the satellite map feature map. Represents the horizontal coordinates of adjacent pixels used to calculate bilinear interpolation, Represents the ordinates of adjacent pixels used to calculate bilinear interpolation, Represents the characteristics of satellite map images, represents the first two-dimensional transformation offset learned for the position, Represents the second two-dimensional transformation offset learned for the position, Indicates taking the maximum value.

6. The lane line detection method combining standard definition map and satellite map fusion according to claim 5, characterized in that: In the step 2, the SD map image is encoded using a polyline sequence representation method to obtain SD map features, which specifically includes the following steps: Segment sampling is performed on the SD map image to obtain a sampled target, and the sampled target is encoded by sinusoidal embedding to obtain an encoding result; Perform one-hot encoding on the lane type to obtain a label representing the road type; The encoding result is concatenated with the one-hot vector representing the road type label to obtain the SD map feature.

7. The lane line detection method combining standard definition map and satellite map fusion according to claim 6, characterized in that: In the step of performing line segment sampling on the SD map image to obtain the sampled target and encoding the sampled target through sine embedding, the corresponding relationship between the process is as follows: ； in, Indicates that after embedding, represents the sine function, represents the cosine function, Indicates the coordinate position, represents the frequency index, represents the influence of temperature scale, represents the temperature scale, represents the scale factor, represents the embedding dimension, Represents division.

8. The lane line detection method combining standard definition map and satellite map fusion according to claim 7, characterized in that: In the step of one-hot encoding the lane type, the corresponding relationship is as follows: ； in, represents the one-hot encoding function, represents the Kroneth function, Indicates the lane category that currently needs to be encoded. represents the total number of lane categories, represents the first category in the lane category set, Indicates the lane category set categories.

9. The lane line detection method combining standard definition map and satellite map fusion according to claim 8, characterized in that: In step 3, the image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing. The corresponding relationship of the process is as follows: ； in, It means that it has been processed by the spatial cross attention mechanism. represents the image BEV feature query sequence, a sequence of keys representing image features, A sequence of values ​​representing image features, represents the scaling numerator; In the step of fusing the processed image BEV features with the SD map features using the multi-head cross attention mechanism, the corresponding relationship between the process is as follows: ； in, It means that it has been processed by the multi-head cross self-attention mechanism. Indicates the corresponding The mapping matrix of the image BEV feature query matrix is: Indicates the corresponding The mapping matrix of the SD map feature key matrix, Indicates the corresponding The mapping matrix of the SD map eigenvalue matrix, represents the weight matrix, Indicates a connection operation. Indicates branches, represents the query sequence of BEV features processed by the spatial cross-attention mechanism, A key sequence representing SD map features, A sequence of values ​​representing SD map features.

10. A lane line detection system combining standard definition map and satellite map, characterized in that: The system applies the lane line detection method combining standard definition map and satellite map fusion as described in any one of claims 1 to 9, and the system comprises: Content input module for: Image acquisition is performed based on the open lane line dataset to obtain multi-view images from vehicle-mounted cameras, SD map images, and satellite map images; SD map feature fusion module, used for: An extraction network is constructed based on the Transformer encoder, and the SD map image is encoded using a polyline sequence representation method to obtain SD map features. Based on the extraction network, the multi-view images of the vehicle camera are input into the image backbone network for feature extraction to obtain image features and image BEV features; The image features and image BEV features are input into the BEV generator for spatial cross self-attention mechanism processing to obtain the processed image BEV features; and the processed image BEV features are fused with the SD map features using the multi-head cross attention mechanism to obtain the BEV features; Satellite map feature fusion module, used for: The U-net network is used to obtain satellite map image features, and the satellite map image features and BEV features are aligned and fused to obtain the final fusion features. Lane topology prediction output module, used for: The final fused features are decoded to generate lane topology prediction results.

Images