A lane line segmentation method and system based on mixed experts

Abstract

The application discloses a lane line segmentation method and system based on a hybrid expert, and belongs to the field of automatic driving and computer vision, aiming at solving the problems that the existing method is difficult to capture long-range dependence of lane lines, has high computational complexity, and a single prediction head is difficult to consider different scales and morphological lane line targets. The application pre-processes a road scene image; a lane line segmentation model based on a hybrid expert network is used to extract multi-scale feature maps through a visual state space model encoder, to enhance the spatial interaction ability of features at each layer through a multi-head hybrid module with multi-head subspace decomposition and a global context broadcasting mechanism, to realize high-quality cross-layer feature fusion through an efficient upsampling module based on pixel rearrangement and channel projection by using a multi-scale pixel decoder; a hybrid expert predictor is used to adaptively fuse feature responses at different stages through a multi-scale expert branch and a shared gating mechanism, and the lane line segmentation result is output by deconvolution to restore the original resolution.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving and computer vision technology, specifically relating to a lane line segmentation method and system based on hybrid experts. Background Technology

[0002] Lane segmentation is one of the core tasks in autonomous driving environmental perception systems. Its goal is to accurately extract lane line regions from road scene images, providing crucial information for vehicle path planning, lane keeping, and lane departure warning. However, lane segmentation in real-world road scenarios faces numerous challenges: lane lines are typically long and thin, with large spans and discrete distributions; lane line boundaries are often blurred or discontinuous due to factors such as changes in lighting, rain and snow occlusion, lane line wear, and vehicle occlusion; furthermore, lane line morphology exhibits significant heterogeneity, with various types including solid lines, dashed lines, and double lines, making it difficult for a single prediction model to simultaneously accommodate lane line targets of different scales and morphologies.

[0003] Traditional lane segmentation methods based on convolutional neural networks have made some progress in local feature extraction by fusing multi-scale features through encoder-decoder structures and skip connections. However, the receptive field of convolutional operations is limited, making it difficult to fully model the long-range dependencies of lane lines, and prone to breakage or missed detection when processing lane lines extending a long distance. To compensate for this deficiency, some methods introduce attention mechanisms or deepen the number of network layers, but this significantly increases the number of model parameters and computational burden, making it difficult to meet the real-time requirements of autonomous driving systems.

[0004] Visual transformers achieve global context modeling through a self-attention mechanism, enabling direct modeling of long-range relationships between pixels and offering advantages in complex structure recognition. However, the computational complexity of self-attention increases quadratically with image resolution, posing extremely high computational challenges in high-resolution road image processing, making it difficult to meet the requirements of embedded deployment and real-time inference. Furthermore, the pure transformer structure lacks local inductive bias, placing higher demands on the training stability of small-sample lane line data.

[0005] Visual state-space models achieve long-range dependency modeling through a linear recursive mechanism, significantly reducing computational complexity while maintaining global perception capabilities, and combining the advantages of convolutional networks and transformers. However, existing methods still have shortcomings in multi-scale feature decoding and adaptive modeling of heterogeneous lane line targets, especially lacking effective mechanisms for handling lane line regions with significant scale changes, blurred boundaries, and heterogeneous morphologies. A single prediction head cannot simultaneously take into account different types of targets. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a lane segmentation method and system based on hybrid experts. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a lane segmentation method based on hybrid experts, comprising: Acquire a road scene image, and preprocess the road scene image to obtain a preprocessed image; The pre-trained lane line segmentation model based on a hybrid expert network is used to perform lane line segmentation on the preprocessed image to obtain lane line segmentation results. The lane segmentation model based on a hybrid expert network performs layer-by-layer encoding feature extraction, spatial interaction enhancement, layer-by-layer upsampling and cross-layer feature fusion, adaptive fusion based on multi-scale expert branches and shared gating mechanisms, and deconvolution and activation function processing on the preprocessed image to achieve lane segmentation.

[0007] In a second aspect, the present invention provides a lane segmentation system based on hybrid experts for performing the lane segmentation method based on hybrid experts as described in the first aspect, the lane segmentation system comprising: The preprocessing module is used to acquire road scene images, preprocess the road scene images, and obtain preprocessed images. A pre-trained lane segmentation model based on a hybrid expert network is used to extract layer-by-layer encoded features from the preprocessed image, resulting in multi-scale feature maps at multiple stages. Spatial interaction enhancement is performed on the multi-scale feature maps at each stage to obtain enhanced features at each stage. Layer-by-layer upsampling and cross-layer feature fusion are then performed on the enhanced features at each stage to obtain multi-scale decoding features at each stage. Based on multi-scale expert branches and a shared gating mechanism, independent expert branches are constructed for feature transformation of the multi-scale decoding features at each stage. The fused features are obtained through feature concatenation and adaptive fusion. The lane segmentation result is obtained based on the fused features.

[0008] The beneficial effects of this invention are: The solution provided in this invention designs a lane line segmentation model based on a hybrid expert network. Utilizing a multi-head hybrid module, it achieves global interaction across spatial locations with extremely low parameter cost through multi-head subspace decomposition and global context broadcasting mechanisms. This eliminates the need to construct pairwise relationship matrices between pixels, and the computational complexity is only linearly related to spatial size, making it suitable for real-time processing of high-resolution road images. A multi-scale pixel decoder, based on pixel rearrangement and channel projection strategies, achieves high-quality upsampling of feature maps, effectively avoiding boundary blurring and parameter redundancy problems caused by traditional bilinear interpolation or transposed convolution, significantly improving the ability to recover lane line edge details. A hybrid expert predictor, through multi-scale expert branching and a shared gating mechanism, adaptively fuses feature responses from different levels, enabling the model to dynamically select appropriate feature representations for lane line regions of different scales, shapes, and boundary clarity, effectively improving the model's segmentation accuracy and robustness for heterogeneous lane line targets. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a lane segmentation method based on hybrid experts provided in an embodiment of the present invention. Figure 2 This is an overall framework diagram of a lane segmentation method based on hybrid experts provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the high-efficiency upsampling module EUM in a lane segmentation method based on hybrid experts provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the shared expert structure in a hybrid expert predictor of a lane segmentation method based on hybrid experts provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the expert structure of the hybrid shared experts in a lane line segmentation method based on hybrid experts provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the lane line segmentation result provided by an embodiment of the present invention, based on a hybrid expert-based lane line segmentation method. Detailed Implementation

[0010] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0011] To address the challenges of convolutional neural networks failing to capture long-range dependencies in lane lines, the excessive computational complexity of transformer models, and the difficulty of a single prediction head taking into account lane line targets of different scales and shapes, this invention provides a lane line segmentation method and system based on hybrid experts.

[0012] Below, we will first introduce a lane segmentation method based on hybrid experts provided in the embodiments of the present invention.

[0013] like Figure 1 As shown in the figure, the lane segmentation method based on hybrid experts provided by the present invention may include the following steps: S1, acquire a road scene image, preprocess the road scene image, and obtain a preprocessed image; S2, use a pre-trained lane line segmentation model based on a hybrid expert network to perform lane line segmentation on the preprocessed image to obtain lane line segmentation results; Among them, the lane line segmentation model based on hybrid expert network performs layer-by-layer encoding feature extraction, spatial interaction enhancement, layer-by-layer upsampling and cross-layer feature fusion, adaptive fusion based on multi-scale expert branch and shared gating mechanism, deconvolution and activation function processing on the preprocessed image to achieve lane line segmentation.

[0014] Preprocessing road scene images to obtain preprocessed images can include: The road scene image is subjected to random horizontal flipping, random vertical flipping, color jittering, and size scaling in sequence to obtain the preprocessed image.

[0015] Lane segmentation models based on hybrid expert networks, such as Figure 2 As shown, it can include the following settings in sequence: The visual state space model encoder is used to perform layer-by-layer encoding feature extraction on the preprocessed image to obtain multi-scale feature maps at multiple stages. A multi-scale pixel decoder is used to perform spatial interactive enhancement on the multi-scale feature maps of each stage to obtain the enhanced features of each stage; the enhanced features of each stage are then upsampled layer by layer and fused with cross-layer features to obtain the multi-scale decoded features of each stage. The hybrid expert predictor is used to construct independent expert branches for feature transformation based on multi-scale expert branches and shared gating mechanism. The fused features are obtained by feature concatenation and adaptive fusion. The lane line segmentation results are obtained by processing the fused features through deconvolution and activation function.

[0016] A lane segmentation model based on a hybrid expert network is constructed using a pre-trained visual state space model encoder, a multi-scale pixel decoder, and a hybrid expert predictor.

[0017] The visual state space model encoder includes multiple coding layers arranged sequentially from top to bottom; the multi-scale pixel decoder includes multiple multi-head mixing modules and upsampling modules. The number of multi-head mixing modules is the same as the number of coding layers. Each multi-head mixing module is set after its corresponding coding layer, and each upsampling module is set after all the multi-head mixing modules except for the multi-head mixing module corresponding to the last stage.

[0018] Preferably, the number of coding layers can be set to four, with each coding layer followed by a multi-head mixing module. An upsampling module is then set after each of the multi-head mixing modules in the first to third stages. The upsampling module is an Enhanced Upscaling Module (EUM).

[0019] The multi-scale pixel decoder restores spatial resolution layer by layer and achieves cross-layer feature fusion through the efficient upsampling module EUM.

[0020] The training process of a lane segmentation model based on a hybrid expert network may include: Road scene images are obtained from the dataset and preprocessed. The preprocessed images are then used to construct a data sample set. The encoder parameters of the visual state space model are adjusted and pre-trained weights are loaded. Encoding features are extracted from images in the data sample set, and multi-scale feature maps of four stages are obtained through four encoding layers. After the multi-scale feature maps of each stage are enhanced by the spatial interaction of the multi-head hybrid module, the high-efficiency upsampling module EUM performs upsampling and cross-layer feature fusion layer by layer to obtain the multi-scale decoding features of the same spatial size for each stage. The multi-scale decoded features are input into the hybrid expert predictor, and an independent expert branch is constructed for the multi-scale decoded features at each stage to obtain the fused features; The fused features are restored to the original image resolution by deconvolution, and the lane line segmentation result is obtained by processing with an activation function. Based on the lane line segmentation results, the parameters and weights of the lane line segmentation model based on the hybrid expert network are optimized through the loss function to obtain a trained lane line segmentation model based on the hybrid expert network for lane line segmentation.

[0021] The loss function designed in this embodiment of the invention adopts binary cross-entropy loss. With Dice loss Joint loss function The specific formula is as follows: ; ; ; in, This indicates the number of pixels in the lane line segmentation result. This indicates the actual result corresponding to the lane line segmentation result. The actual pixel value of the location. In the lane line segmentation results The predicted pixel value for the location. Both losses are weighted at 0.5; by gradually decreasing the loss, the model weights are optimized.

[0022] The process of enhancing spatial interaction through multi-head hybrid modules may include: The multi-scale feature map of each stage is divided into K subspaces along the channel dimension; K is an integer greater than 0. Spatial blending is performed independently for each subspace, and the global context of each subspace is obtained through global average pooling; By using convolutional projection, the dimension of the global context is transformed to the size of the original subspace, and the output features of each subspace are obtained. The output features of all subspaces are concatenated and normalized to obtain the enhanced features of each stage.

[0023] This invention enhances the spatial interaction capability of multi-scale feature maps at various stages with low parameter cost through a multi-head hybrid module. The multi-head hybrid module, through multi-head subspace decomposition and global context broadcasting, achieves global interaction across spatial locations with extremely low parameter cost. It eliminates the need to construct pairwise relationship matrices between pixels, and its computational complexity is only linearly related to spatial size, making it suitable for real-time processing of high-resolution road images.

[0024] The upsampling module's process of performing layer-by-layer upsampling and cross-layer feature fusion can include: Depth-separable convolution is used to perform channel expansion on the enhanced features at each stage to obtain channel-expanded features. Upsampling of the channel-expanded features is performed by using pixel rearrangement, rearranging the information from the channel dimension to the spatial dimension to obtain the enlarged features. The enlarged features are used to generate intermediate features through convolution, which are then used for cross-layer residual fusion in the next stage. Projecting the amplified features after cyclic upsampling onto a unified dimension yields multi-scale decoding features for each stage. Rearranging the channel-dimensional information into the spatial dimension can double the length and width of the amplified features.

[0025] The working process of each upsampling module, such as Figure 2As shown, each of the multi-head mixing modules in the first, second, and third stages is followed by an upsampling module, which is a high-efficiency upsampling module (EUM). Each EUM processes the enhanced features of its corresponding stage, outputting intermediate features to the next stage for cross-layer residual fusion. The multi-scale decoded features are then output to the hybrid expert predictor. For the EUM corresponding to the first stage, its input is the output of the multi-head mixing module corresponding to the first stage, and its output is the intermediate features of the first stage. and multi-scale decoding features For the high-efficiency upsampling module EUM corresponding to the second stage, its input is the intermediate feature between the output of the multi-head mixing module corresponding to the second stage and the output of the high-efficiency upsampling module EUM corresponding to the first stage. The result after cross-layer residual fusion is the output of the intermediate features corresponding to the second stage. and multi-scale decoding features For the high-efficiency upsampling module EUM corresponding to the third stage, its input is the intermediate feature between the output of the multi-head mixing module corresponding to the third stage and the output of the high-efficiency upsampling module EUM corresponding to the second stage. The result after cross-layer residual fusion is the output of the intermediate features corresponding to the third stage. and multi-scale decoding features .

[0026] A schematic diagram of the high-efficiency upsampling module EUM is shown below. Figure 3 As shown, it can include several upsampling blocks arranged in sequence, with 3×3 convolutions set after the first and last upsampling blocks respectively; each upsampling block has a depthwise convolution and a normalization layer arranged in sequence.

[0027] The embodiments of the present invention design an efficient upsampling module, which achieves high-quality upsampling of feature maps based on pixel rearrangement and channel projection strategies. This effectively avoids the boundary blurring and parameter redundancy problems caused by traditional bilinear interpolation or transposed convolution, and significantly improves the ability to restore lane line edge details.

[0028] The process of adaptive fusion of multi-scale expert branching and shared gating mechanisms in a hybrid expert predictor may include: For the multi-scale decoding features of each stage, a corresponding independent expert branch is constructed, and the feature transformation is performed in the form of gated linear unit activation to obtain the independent expert branch features corresponding to each stage. Shared features are generated by concatenating the multi-scale decoding features corresponding to all stages using shared experts, and the weights corresponding to each stage are calculated using a normalized exponential function. The features of the independent expert branches corresponding to each stage are weighted and summed according to the weights of each stage, and then added to the corresponding shared features to obtain the fused features.

[0029] For the multi-scale decoding features at each stage, an independent expert branch is constructed to perform feature transformation.

[0030] Feature transformation is performed using gated linear unit activation. This includes: Gated mapping and nonlinear transformation are performed on the multi-scale decoding features of each stage. The mapping results and transformation results are multiplied element by element and then linearly projected to obtain the independent expert branch features corresponding to each stage.

[0031] A structural diagram of shared experts, such as Figure 4 As shown, in the shared expert model, the output of the 3×3 convolution is processed by both the expert and the 3×3 convolution. The data processed by the 3×3 convolution is then processed by the Softmax function. The outputs processed by the expert and the outputs processed by the Softmax function are used as the weights for each stage. The structure of the expert is illustrated in the diagram below. Figure 5 As shown, parallel upsampling blocks and gated convolutions are set up in the expert. The output of the gated convolution is added to the output of the upsampling block after passing through the Swish function, and then passed through the downsampling block to obtain the corresponding output of the expert.

[0032] The process of obtaining lane line segmentation results using a segmentation head based on fusion features can include: The segmentation head in the hybrid expert predictor uses deconvolution to progressively enlarge the spatial resolution of the fused features to match that of the road scene image, and then scales it using an activation function to obtain the lane line segmentation result.

[0033] The present invention designs a hybrid expert predictor, which adaptively fuses feature responses from different stages through multi-scale expert branching and shared gating mechanism. This enables the model to dynamically select appropriate feature representations for lane line regions of different scales, shapes and boundary clarity, effectively improving the model's segmentation accuracy and robustness for heterogeneous lane line targets.

[0034] Lane segmentation results, such as Figure 6 As shown, the lane line segmentation results output by the lane line segmentation model based on a hybrid expert network provided in this embodiment of the invention are clear and complete, with no omissions.

[0035] The lane segmentation method based on hybrid experts provided in this invention performs data augmentation and preprocessing on the acquired road scene image. Then, it extracts multi-scale feature maps using a pre-trained visual state space model encoder in the lane segmentation model based on a hybrid expert network. A multi-head fusion module enhances the spatial interaction capability of the multi-scale feature maps at each stage with low parameter cost. Next, a multi-scale pixel decoder restores the spatial resolution stage by stage, and a high-quality cross-layer feature fusion is achieved through an efficient upsampling module. Finally, a hybrid expert predictor adaptively fuses features from different stages through multi-scale expert branches and a shared gating mechanism to output the lane segmentation result. This invention achieves high-precision lane segmentation with a small number of trainable parameters, effectively improving the model's adaptability to lane lines with large scale variations, blurred boundaries, and heterogeneous morphologies.

[0036] Secondly, corresponding to the above method embodiments, this invention also provides a lane segmentation system based on hybrid experts, used to execute the lane segmentation method based on hybrid experts as described in the first aspect. The lane segmentation system may include: The preprocessing module is used to acquire road scene images, preprocess the road scene images, and obtain preprocessed images. A pre-trained lane segmentation model based on a hybrid expert network is used to extract layer-by-layer encoded features from the preprocessed image, resulting in multi-scale feature maps at multiple stages. Spatial interaction enhancement is performed on the multi-scale feature maps at each stage to obtain enhanced features at each stage. Layer-by-layer upsampling and cross-layer feature fusion are then performed on the enhanced features at each stage to obtain multi-scale decoding features at each stage. Based on multi-scale expert branches and a shared gating mechanism, independent expert branches are constructed for feature transformation of the multi-scale decoding features at each stage. The fused features are obtained through feature concatenation and adaptive fusion. The lane segmentation result is obtained based on the fused features.

[0037] The beneficial effects of the embodiments of the present invention will be further explained below with reference to simulation experiments.

[0038] Simulation conditions: The lane segmentation model based on a hybrid expert network constructed in this embodiment of the invention is implemented using a deep learning framework and trained on a high-performance graphics processor. All input road scene images are uniformly scaled to a resolution of 256×256. During training, data augmentation strategies such as random horizontal flipping, random vertical flipping, and color dithering are employed. An optimizer combined with a cosine annealing learning rate scheduling strategy is used, with an initial learning rate set to 2e-4, a batch size of 8, and a maximum number of training epochs of 300.

[0039] Simulation content: The effectiveness of the embodiments of the present invention is verified using a publicly available lane line segmentation dataset. The lane line segmentation dataset contains lane line annotation images under various road scenarios, covering daytime, nighttime, rainy / snowy weather, and different lighting conditions. Lane line types include solid lines, dashed lines, and double lines. For performance evaluation, the intersection-over-union ratio (IoU) and Dice coefficient are used uniformly as evaluation metrics. Simultaneously, the number of model parameters and computational complexity are compared to measure the accuracy and practicality of the method.

[0040] Simulation Results: Experimental results show that the embodiments of the present invention outperform existing state-of-the-art methods in terms of intersection-over-union ratio and Dice coefficient, with only about 10.68M parameters and a computational load as low as 1.11 billion floating-point operations. The embodiments of the present invention dynamically activate feature channels adapted to the current input through a hybrid expert mechanism, enhancing the model's ability to express irregularly shaped lane lines and subtle textures. Simultaneously, leveraging its sparsely activated expert network, it can adaptively fuse multi-scale features for different local regions, effectively suppressing noise interference and strengthening weak boundary responses, thereby achieving superior lane line segmentation performance in difficult-to-identify scenarios.

[0041] This invention utilizes a multi-head hybrid module, employing multi-head subspace decomposition and global context broadcasting mechanisms to achieve global interaction across spatial locations with extremely low parameter costs. It eliminates the need to construct pairwise relationship matrices between pixels, and its computational complexity is only linearly related to spatial size, making it suitable for real-time processing of high-resolution road images. A multi-scale pixel decoder, based on pixel rearrangement and channel projection strategies, achieves high-quality upsampling of feature maps, effectively avoiding boundary blurring and parameter redundancy issues caused by traditional bilinear interpolation or transposed convolution, significantly improving the recovery of lane line edge details. A hybrid expert predictor, through multi-scale expert branching and a shared gating mechanism, adaptively fuses feature responses from different levels, enabling the model to dynamically select appropriate feature representations for lane line regions of different scales, shapes, and boundary clarity, effectively improving the model's segmentation accuracy and robustness against heterogeneous lane line targets. This invention achieves high-precision lane line segmentation with a small number of trainable parameters, effectively improving the adaptability of the hybrid expert network-based lane line segmentation model to lane lines with large scale variations, blurred boundaries, and heterogeneous shapes, making it suitable for autonomous driving environment perception in complex road scenarios.

[0042] It should be noted that, in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0043] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A lane segmentation method based on hybrid experts, characterized in that, include: Acquire a road scene image, and preprocess the road scene image to obtain a preprocessed image; The pre-trained lane line segmentation model based on a hybrid expert network is used to perform lane line segmentation on the preprocessed image to obtain lane line segmentation results. The lane segmentation model based on a hybrid expert network performs layer-by-layer encoding feature extraction, spatial interaction enhancement, layer-by-layer upsampling and cross-layer feature fusion, adaptive fusion based on multi-scale expert branches and shared gating mechanisms, and deconvolution and activation function processing on the preprocessed image to achieve lane segmentation.

2. The lane segmentation method based on hybrid experts according to claim 1, characterized in that, The road scene image is preprocessed to obtain a preprocessed image, including: The road scene image is sequentially subjected to random horizontal flipping, random vertical flipping, color jittering, and size scaling to obtain a preprocessed image.

3. The lane segmentation method based on hybrid experts according to claim 1, characterized in that, The lane segmentation model based on a hybrid expert network includes the following components set sequentially: The visual state space model encoder is used to perform layer-by-layer encoding feature extraction on the preprocessed image to obtain multi-scale feature maps at multiple stages. A multi-scale pixel decoder is used to perform spatial interactive enhancement on the multi-scale feature maps of each stage to obtain the enhanced features of each stage. The enhanced features at each stage are upsampled layer by layer and fused with cross-layer features to obtain multi-scale decoding features for each stage. A hybrid expert predictor is used to construct independent expert branches for feature transformation based on multi-scale expert branching and shared gating mechanism, and obtain fused features through feature concatenation and adaptive fusion. Lane segmentation results are obtained by processing the fusion features through deconvolution and activation functions.

4. The lane segmentation method based on hybrid experts according to claim 3, characterized in that, The visual state space model encoder includes multiple coding layers arranged sequentially from top to bottom; the multi-scale pixel decoder includes multiple multi-head mixing modules and upsampling modules. The number of multi-head mixing modules is the same as the number of coding layers. Each multi-head mixing module is set after its corresponding coding layer, and each upsampling module is set after all the other multi-head mixing modules except for the multi-head mixing module corresponding to the last stage.

5. The lane segmentation method based on hybrid experts according to claim 4, characterized in that, The process of enhancing spatial interaction by the multi-head hybrid module includes: The multi-scale feature map of each stage is divided into K subspaces along the channel dimension; K is an integer greater than 0. Spatial blending is performed independently for each subspace, and the global context of each subspace is obtained through global average pooling; By using convolutional projection, the dimension of the global context is transformed to the size of the original subspace, and the output features of each subspace are obtained. The output features of all subspaces are concatenated and normalized to obtain the enhanced features of each stage.

6. The lane segmentation method based on hybrid experts according to claim 4, characterized in that, The upsampling module performs layer-by-layer upsampling and cross-layer feature fusion processes, including: Depth-separable convolution is used to perform channel expansion on the enhanced features at each stage to obtain channel-expanded features. Upsampling of the channel-expanded features is performed by using pixel rearrangement, rearranging the information from the channel dimension to the spatial dimension to obtain the enlarged features. The enlarged features are used to generate intermediate features through convolution, which are then used for cross-layer residual fusion in the next stage. The enlarged features after cyclic upsampling are projected onto a unified dimension to obtain multi-scale decoding features at each stage.

7. The lane segmentation method based on hybrid experts according to claim 3, characterized in that, The process of adaptive fusion of the hybrid expert predictor based on multi-scale expert branching and shared gating mechanism includes: For the multi-scale decoding features of each stage, a corresponding independent expert branch is constructed, and the feature transformation is performed in the form of gated linear unit activation to obtain the independent expert branch features corresponding to each stage. The multi-scale decoding features corresponding to all stages are concatenated to generate shared features, and the weights corresponding to each stage are calculated using a normalized exponential function. The features of the independent expert branches corresponding to each stage are weighted and summed according to the weights of each stage, and then added to the corresponding shared features to obtain the fused features.

8. The lane segmentation method based on hybrid experts according to claim 7, characterized in that, Feature transformation is performed using gated linear unit activation, including: Gated mapping and nonlinear transformation are performed on the multi-scale decoding features of each stage. The mapping results and transformation results are multiplied element by element and then linearly projected to obtain the independent expert branch features corresponding to each stage.

9. The lane segmentation method based on hybrid experts according to claim 3, characterized in that, The process of obtaining lane line segmentation results by processing fused features through deconvolution and activation functions includes: The segmentation head in the hybrid expert predictor uses deconvolution to progressively enlarge the spatial resolution of the fused features to the same level as the road scene image, and then scales it using an activation function to obtain the lane line segmentation result.

10. A lane segmentation system based on hybrid experts, for executing the lane segmentation method based on hybrid experts as described in any one of claims 1-9, wherein the lane segmentation system is characterized in that, include: The preprocessing module is used to acquire road scene images, preprocess the road scene images, and obtain preprocessed images. A pre-trained lane segmentation model based on a hybrid expert network is used to extract layer-by-layer encoded features from the preprocessed image, resulting in multi-scale feature maps at multiple stages. Spatial interaction enhancement is then performed on the multi-scale feature maps at each stage to obtain the enhanced features at each stage. The enhanced features at each stage are upsampled layer by layer and fused with cross-layer features to obtain multi-scale decoding features for each stage. Based on multi-scale expert branching and shared gating mechanism, independent expert branches are constructed for feature transformation of multi-scale decoding features at each stage, and fused features are obtained through feature splicing and adaptive fusion. Lane segmentation results are obtained based on fusion features.

Images