Lane detection method based on global feature optimization and shape feature perception aggregation

Abstract

This invention relates to a lane detection method based on global feature optimization and shape feature perception aggregation, belonging to the field of autonomous driving technology. The method proposes a lane detection model that integrates two core components: a global feature optimizer (GFO) and an adaptive lane geometry aggregator (ALGA). The global feature optimizer enables the model to fully fuse local and global features in the image, effectively filtering out unimportant information. Simultaneously, the adaptive lane geometry aggregator effectively utilizes the unique shape prior information of the lane lines, allowing the model to effectively handle complex scenarios such as curves and missing lane lines, effectively mitigating the performance degradation problem of previous methods when encountering complex situations.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a lane line detection method based on global feature optimization and shape feature perception aggregation. Background Technology

[0002] Currently, autonomous driving [1] technology has attracted great attention in both academia and industry. Among them, lane detection, as a core component of intelligent vehicle systems, plays a crucial role. Its main task is to accurately identify and locate lane lines from various road environments, which is essential for ensuring vehicle driving safety and route planning.

[0003] Existing lane detection methods can be divided into two categories: traditional image processing-based methods and deep learning-based methods. Image processing-based methods mainly rely on traditional image processing techniques such as edge detection [2] and Hough transform [3]. These methods identify lane lines by detecting edge features in the image, but because they rely on manually designed features, they are easily affected by external factors such as changes in lighting, shadows, and road wear, resulting in low detection accuracy and stability. For example, the common Canny edge detection algorithm and Hough transform method can provide certain detection results under ideal conditions, but they perform poorly in actual complex environments.

[0004] With the development of deep learning technology, lane detection methods based on convolutional neural networks (CNN) have become a research hotspot. Deep learning-based methods are mainly divided into four categories: segmentation-based methods, curve-based methods, keypoint-based methods, and anchor-based methods. Segmentation-based methods regard lane detection as a pixel-by-pixel segmentation problem, representing each pixel as a lane line region or background region. For example, the SCNN[4] (Spatial CNN) method transmits information in the rows and columns of the image through a slice network structure, achieving a high detection accuracy. However, these methods usually require a long processing time and have a large computational load, making them unsuitable for real-time applications. Curve-based methods[5] model the lane curve and then regress these parameters to detect the lane. For example, the PolyLaneNet[6] method uses polynomial coefficients to represent the curve, achieving high efficiency. BezierLaneNet[7] proposes to use parameterized Bezier curves to achieve end-to-end lane detection. However, since the polynomial coefficients of the curve abstraction are not easy to learn, the detection performance of these methods is low. Keypoint-based methods detect lane points as keypoints[8] and then group them into lane instances. For example, the PINet[9] method predicts key point locations and feature embeddings through stacked hourglass networks, and then clusters the key points into lane instances. Such methods require complex post-processing and have high computational costs. Anchor-based methods use predefined line anchors as references to regress accurate lanes. For example, Line-CNN

[11] uses straight lines with specific directions emitted from the image boundary as anchors, and the CLRNet

[10] (Cross Layer Refinement Network) method uses cross-layer refinement networks to detect the approximate location of lanes using high-level semantic features, and then refines layer by layer based on low-level features, while always maintaining the high-level semantic information of the lane line image throughout the process. This method achieves a good balance between detection accuracy and speed, but its performance in complex scenes still needs improvement.

[0005] Existing lane detection methods often achieve high accuracy in routine scenarios. However, their accuracy drops significantly in complex road conditions (such as curves, lane occlusion and missing lane lines, and strong light). This is because these methods fail to effectively distinguish between important lane line information and redundant background information during feature extraction, leading to wasted computational resources and low detection efficiency. Furthermore, when processing global and local lane line features, the models struggle to effectively fuse them, failing to fully utilize prior knowledge of lane line shape and global information in the image. This results in poor lane line detection performance in complex situations, making them susceptible to environmental changes and unable to accurately identify lane lines in dynamic environments.

[0006] References:

[0007] [1]Ekim Yurtsever,Jacob Lambert,Alexander Carballo,and KazuyaTakeda.A survey of autonomous driving:Common practices and emergingtechnologies.IEEE access,8:58443–58469,2020.

[0008] [2]Shou-Ming Hou,Chao-Lan Jia,Ya-Bing Wanga,and Mackenzie Brown.Areview ofthe edge detection technology.Sparklinglight Transactions onArtificial Intelligence and Quantum Computing(STAIQC),1(2):26–37,2021.

[0009] [3]Kai Zhao,Qi Han,Chang-Bin Zhang,Jun Xu,and Ming-Ming Cheng.Deephough transform for semantic line detection.IEEE Transactions on PatternAnalysis and Machine Intelligence,44(9):4793–4806,2021.

[0010] [4]Xingang Pan,Jianping Shi,Ping Luo,Xiaogang Wang,and XiaoouTang.Spatial as deep:Spatial cnn for traffic scene understanding.InProceedings of the AAAI Conference on Artificial Intelligence,volume 32,2018.

[0011] [5]Wouter Van Gansbeke,Bert De Brabandere,Davy Neven,Marc Proesmans,and Luc Van Gool.End-to-end lane detection through differentiable leastsquares fitting.In Proceedings ofthe IEEE / CVF International Conference onComputer Vision Workshops,pages 0–0,2019.

[0012] [6]Lucas Tabelini,Rodrigo Berriel,Thiago M Paixao,Claudine Badue,Alberto F De Souza,and Thiago Oliveira-Santos.Polylanenet:Lane estimation viadeep polynomial regression.In 2020 25th International Conference on PatternRecognition(ICPR),pages 6150–6156.IEEE,2021.

[0013] [7]Zhengyang Feng,Shaohua Guo,Xin Tan,Ke Xu,Min Wang,and LizhuangMa.Rethinking efficient lane detection via curve modeling.In Proceedingsofthe IEEE / CVF Conference on Computer Vision and Pattern Recognition,pages17062–17070,2022.

[0014] [8]Ce Zheng, Wenhan Wu, Chen Chen, Taojiannan Yang, Sijie Zhu, Ju Shen, Nasser Kehtarnavaz, and Mubarak Shah. Deep learning-based human pose estimation: A survey. ACM Computing Surveys, 56(1): 1–37, 2023.

[0015] [9]Yeongmin Ko, Younkwan Lee, Shoaib Azam, Farzeen Munir, Moongu Jeon, and Witold Pedrycz. Key points estimation and point instance segmentation approach for lane detection. IEEE Transactions on Intelligent Transportation Systems, 23(7): 8949–8958, 2021.

[0016]

[10] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, pages 898–907, 2022.

[0017]

[11] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1): 248–258, 2019. Summary of the Invention

[0018] The purpose of this invention is to address the problems existing in the background technology by providing a lane line detection method based on global feature optimization and shape feature perception aggregation, which can achieve high robustness and high accuracy of lane line detection in complex road environments.

[0019] To achieve the above objectives, the technical solution of this invention is: a lane detection method based on global feature optimization and shape feature perception aggregation. It proposes a lane detection model that integrates two core components: a global feature optimizer (GFO) and an adaptive lane geometry aggregator (ALGA). The GFO enables the model to fully fuse local and global features in the image, while the ALGA effectively utilizes the prior shape information of the lane lines, thereby achieving lane detection.

[0020] In one embodiment of the present invention, the method includes the following steps:

[0021] Preprocess the input lane line image;

[0022] Features at different levels are extracted using the backbone network;

[0023] The extracted features are processed under the operation of the Global Feature Optimizer (GFO).

[0024] The adaptive lane geometry aggregator ALGA utilizes prior information about the unique shape of lane lines to enhance the relevance of lane anchor points in the image.

[0025] The extracted features are refined layer by layer from coarse to fine through the neck network to optimize the extracted features and maintain high-level semantic information throughout the process.

[0026] The optimized features are processed through operations including pooling, convolution, and fully connected layers, and then fed into the regression and classification branches for feature aggregation and visualization to obtain the final lane line prediction image.

[0027] In one embodiment of the present invention, the preprocessing method for the input lane line image is to crop the lane line image to a size of 320×800.

[0028] In one embodiment of the present invention, the backbone network is ResNet-18.

[0029] In one embodiment of the present invention, the neck network is a feature pyramid.

[0030] In one embodiment of the present invention, the Global Feature Optimizer (GFO) combines the efficient channel attention of ECA and the spatial attention of CBAM to more finely adjust the network’s attention to different channels and spatial regions, enabling the model to better handle complex and variable road environments. At the same time, since the one-dimensional convolution of ECA is lighter than the fully connected layer of CBAM, it can effectively improve the speed of the model. For lane detection systems that require real-time processing, the Global Feature Optimizer can provide fast and accurate performance.

[0031] In one embodiment of the present invention, the Global Feature Optimizer (GFO) is implemented as follows:

[0032] For an input feature map of size H×W×C (where H represents the height of the input feature map, W represents the width of the input feature map, and C represents the number of channels in the input feature map), the first step is to process the efficient channel attention in ECA. This first performs global feature pooling on the input feature map, resulting in a 1×1×C output. Next, a one-dimensional convolution operation is performed on this output to capture the inter-channel dependencies. The size of the convolution kernel is adaptively adjusted according to the number of channels in the input feature map, as shown in the following formula:

[0033]

[0034] Where |t| odd Let represent the nearest odd number to t, and set γ and b as constants;

[0035] The output is then passed through a Sigmoid activation function to obtain the final channel attention features, which are then multiplied element-wise with the original feature map. This process is represented by the following formula:

[0036] F eca =F×σ(Conv 1D (AvgPool(F)))

[0037] AvgPool is a global average pooling method that compresses the spatial dimension of the feature map F, retaining only channel information; Conv... 1D σ is a one-dimensional convolution used to capture the dependencies between channels; σ is the Sigmoid activation function used to generate channel attention weights.

[0038] Next, the feature map of size H×W×C output by ECA continues to be used as input for the spatial attention of CBAM. First, max pooling and average pooling operations are performed on the feature map in the channel dimension to obtain two H×W×1 feature maps. Then, a concat operation is performed on these two feature maps in the channel dimension. Next, a convolution operation with a kernel size of 7×7 is used to reduce the channel dimension and learn the importance of different spatial locations. Then, the sigmoid function is used to generate spatial attention features, and this feature is then multiplied element-wise with the feature map output by ECA to obtain the final feature map that fuses channel and spatial attention. This process is represented by the following formula:

[0039] F spatial =σ(Conv 2D (Concat(AvgPool(F eca ),MaxPool(F eca ))))×F eca

[0040] Among them, AvgPool(F eca ) and MaxPool(F eca ) represent the feature maps F after ECA adjustment. eca Perform average pooling and max pooling; Concat concatenates the results of the two pooling operations along the channel dimension; Conv 2D σ is a two-dimensional convolution operation used to learn the importance of different spatial locations; σ is the sigmoid activation function used to generate spatial attention weights.

[0041] In one embodiment of the present invention, γ and b are set to 2 and 1, respectively.

[0042] In one embodiment of the present invention, the adaptive lane geometry aggregator (ALGA) is used to perceive the shape features of lane lines. It predicts the offset between anchor points and their surrounding anchor points and the tuning factor of each offset, and uses deformable convolution to combine these offsets and tuning factors to integrate features and aggregate the contextual information of anchor points.

[0043] In one embodiment of the present invention, the adaptive lane geometry aggregator (ALGA) is specifically implemented as follows:

[0044] For a lane anchor point, a convolutional layer is first used to predict its offset from K surrounding anchor points in the same lane and the corresponding tuning factor for each offset. This tuning factor allows the model to adaptively adjust the contribution of each anchor point, as shown in the following formula:

[0045]

[0046] Where, pi This represents the coordinates of the i-th anchor point. Let represent the feature representation of the i-th anchor point, φ represent the offset convolutional layer used to learn and extract the predicted offset from the feature map, and ψ represent the quality adjustment factor convolutional layer used to learn and extract the quality adjustment factor for each offset from the feature map. This represents the set of predicted offsets. This represents the offset of the m-th offset. This represents the set of conditioning factors indicating the predicted offset. This represents the conditioning factor for the m-th offset;

[0047] Next, deformable convolution is used to integrate the features adjacent to the anchor point by combining the predicted offset and the quality adjustment factor corresponding to each offset, so as to aggregate the context information of the i-th anchor point, as shown below:

[0048]

[0049] Among them, w m m = 1, ..., M are the weights of the convolution;

[0050] After feature aggregation through deformable convolution, spatial attention is applied to the feature map to effectively handle cases of missing or occluded lane lines, resulting in the final lane line prediction image.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The innovation of this invention lies in (1) proposing a multi-scale attention mechanism that can simulate the human visual focusing ability, called the global feature optimizer, which can effectively filter out unimportant information and focus on processing the image regions most relevant to the task, alleviating the problem of insufficient lane image feature discrimination caused by redundant background information, while reducing computational cost and improving processing efficiency and accuracy without sacrificing performance; (2) designing a shape feature perception module called the adaptive lane geometry aggregator, which can utilize the special shape prior of lane lines to enhance the correlation of anchor points in the image, thereby improving the model's ability to process details of slender and curved lane lines in complex road environments, especially in the presence of occlusion, changes in lighting, or poor weather conditions; (3) by integrating the global feature optimizer and the adaptive lane geometry aggregator into the existing model, the model can exhibit higher performance stability in complex road scenarios, and a large number of experiments and visualization of experimental results have been conducted to effectively verify that the algorithm model of this invention can perform lane line detection better and has better lane line detection performance than the benchmark model in complex scenarios, especially in curved scenarios, it can improve the F1 score by nearly three percentage points compared to the benchmark model. Attached Figure Description

[0053] Figure 1 This is the network model architecture of the method of the present invention.

[0054] Figure 2 This is a framework diagram of the Global Feature Optimizer (GFO) of this invention.

[0055] Figure 3 This is a block diagram of the Adaptive Lane Geometry Aggregator (ALGA) of the present invention.

[0056] Figure 4 Visualize the model's detection results in different scenarios on the CULane dataset. Detailed Implementation

[0057] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0058] This invention proposes a lane detection method based on global feature optimization and shape feature perception aggregation. It proposes a lane detection model that integrates two core components: a global feature optimizer (GFO) and an adaptive lane geometry aggregator (ALGA). The global feature optimizer (GFO) enables the model to fully fuse local and global features in the image, while the adaptive lane geometry aggregator (ALGA) effectively utilizes the prior shape information of the lane lines, thereby achieving lane detection.

[0059] The following is a detailed implementation process of the present invention.

[0060] like Figure 1 The diagram shows the network model architecture of the method of this invention. The method first preprocesses the input lane line image, cropping it to 320×800 pixels. Then, a backbone network (such as ResNet-18) extracts features at different levels. Next, the extracted features are processed by a Global Feature Optimizer (GFO), effectively filtering out unimportant information by simulating human visual focusing, enabling the model to process the lane line regions most relevant to the task. Then, an Adaptive Lane Geometry Aggregator (ALGA) utilizes the special shape prior information of the lane lines to enhance the relevance of lane anchor points in the image. Next, a neck network (such as a feature pyramid) refines the extracted features layer by layer from coarse to fine to optimize the extracted features while maintaining high-level semantic information throughout the process. Finally, the optimized features are subjected to pooling, convolution, and fully connected layers, and then fed into the regression and classification branches for feature aggregation and visualization to obtain the final lane line prediction image. The specific implementation steps of the Global Feature Optimizer (GFO) and the Adaptive Lane Geometry Aggregator (ALGA) are as follows:

[0061] (1) Global Feature Optimizer (GFO)

[0062] To better utilize the important feature information in lane line images and alleviate the problems of background information redundancy and insufficient feature discrimination in lane line regions in previous works, this invention constructs a global feature optimizer that simulates human visual focusing ability for lane line detection. This optimizer can more finely adjust the model's focus on different channels and spatial regions, improving the model's feature representation ability. The global feature optimizer proposed in this invention is as follows: Figure 2 As shown, by combining the efficient channel attention of ECA and the spatial attention of CBAM, the network's attention to different channels and spatial regions can be adjusted more finely, enabling the model to better handle complex and variable road environments, such as different lighting conditions and complex road conditions. Furthermore, since the one-dimensional convolution of ECA is lighter than the fully connected layers of CBAM, it can effectively improve the model's speed. For lane detection systems requiring real-time processing, the global feature optimizer can provide fast and accurate performance. For an input feature map of size H×W×C (where H represents the height of the input feature map, W represents the width of the input feature map, and C represents the number of channels in the input feature map), it first enters the efficient channel attention processing of ECA, which first performs global feature pooling on the input feature map to obtain a 1×1×C output. Next, a one-dimensional convolution operation is performed on this output to capture the inter-channel dependency information. The size of the convolution kernel can be adaptively adjusted according to the number of channels in the input feature map, as shown in the following formula.

[0063]

[0064] Where |t| odd Let represent the nearest odd number to t. In this paper, γ and b are set to 2 and 1, respectively.

[0065] The output is then passed through a Sigmoid activation function to obtain the final channel attention features, which are then multiplied element-wise with the original feature map. This process can be represented by the following formula:

[0066] F eca =F×σ(Conv 1D (AvgPool(F)))

[0067] AvgPool is a global average pooling method that compresses the spatial dimension of the feature map F, retaining only channel information; Conv... 1D σ is a one-dimensional convolution used to capture the dependencies between channels; σ is the Sigmoid activation function used to generate channel attention weights.

[0068] Next, the feature map of size H×W×C output by the efficient channel attention module continues to serve as input for the spatial attention of CBAM. First, max pooling and average pooling operations are performed on the feature map in the channel dimension to obtain two H×W×1 feature maps. Then, a concat operation is performed on these two feature maps in the channel dimension. Next, a convolution operation with a 7×7 kernel is used to reduce the channel dimension and learn the importance of different spatial locations. Then, the sigmoid function is used to generate spatial attention features, and this feature is then multiplied element-wise with the feature map output by ECA to obtain the final feature map that fuses channel and spatial attention. This process can be represented by the following formula:

[0069] F spatial =σ(Conv 2D (Concat(AvgPool(F eca ),MaxPool(F eca ))))×F eca

[0070] Among them, AvgPool(F eca ) and MaxPool(F eca ) represent the feature maps F after ECA adjustment. eca Perform average pooling and max pooling; Concat concatenates the results of the two pooling operations along the channel dimension; Conv 2D σ is a two-dimensional convolution operation used to learn the importance of different spatial locations; σ is the sigmoid activation function used to generate spatial attention weights.

[0071] (2) Adaptive Lane Geometry Aggregator (ALGA)

[0072] To make reasonable use of the special shape prior of lane lines and avoid the limitations of traditional convolutional kernels that sample features within a fixed area and cannot effectively handle elongated lane lines, this invention proposes an adaptive lane geometry aggregator to perceive the shape features of lane lines. This aggregator predicts the offset between the anchor point and its surrounding anchor points, as well as the tuning factor of each offset, and uses deformable convolution to combine these offsets and tuning factors for feature integration, thereby aggregating the contextual information of the anchor points.

[0073] The adaptive lane geometry aggregator proposed in this invention, such as Figure 3 As shown, taking a specific lane anchor point as an example, this invention first uses a convolutional layer to predict the offset between it and K anchor points in the same lane around it, as well as the corresponding tuning factor for each offset. This tuning factor allows the model to adaptively adjust the contribution of each anchor point, as shown in the following formula:

[0074]

[0075] Where, p i This represents the coordinates of the i-th anchor point. Let represent the feature representation of the i-th anchor point, φ represent the offset convolutional layer used to learn and extract the predicted offset from the feature map, and ψ represent the quality adjustment factor convolutional layer used to learn and extract the quality adjustment factor for each offset from the feature map. This represents the set of predicted offsets. This represents the offset of the m-th offset. This represents the set of conditioning factors indicating the predicted offset. This represents the conditioning factor for the m-th offset;

[0076] Next, deformable convolution is used to integrate the features adjacent to the anchor point by combining the predicted offset and the quality adjustment factor corresponding to each offset, so as to aggregate the context information of the i-th anchor point, as shown below:

[0077]

[0078] Among them, w m m = 1, ..., M are the weights of the convolution;

[0079] After feature aggregation via deformable convolution, the model has acquired sufficient local information from the image, allowing it to effectively model the lane line shape using prior knowledge of the lane line's unique shape. Next, spatial attention (the same spatial attention used in the global feature aggregator) is applied to the feature map to further integrate and refine the obtained local features. This enables the model to utilize global information about the image location, fully fusing local and global information to effectively handle situations where lane lines are missing or occluded.

[0080] Implementation Examples

[0081] To evaluate the performance of the lane detection model in complex scenes, this invention was validated on the CULane and TuSimple datasets. CULane is a challenging large-scale dataset for lane detection. It contains 133,235 images, divided into three subsets: training, validation, and test sets. All images are formatted to a size of 1640×590 pixels. The test set is divided into nine groups, eight of which are considered challenging scenarios such as nighttime, bright light, and curves. TuSimple is a widely used benchmark for lane detection. It only contains highway scenes, with most images in clear weather conditions. 3626 images were used for training, 358 for validation, and 2782 for testing. All images are 1280×720 pixels in size.

[0082] Regarding the evaluation metric, for the CULane dataset, this invention uses the F1-Score as the evaluation metric. This invention uses a 30-pixel-wide segmentation mask to calculate the Intersection over Union (IoU) matrix for each prediction-ground (GT) pair. First, this invention filters out predictions with confidence exceeding a predefined threshold, and then selects predictions with segmentation IoU values ​​greater than the threshold t. IoU The predicted-true value (GT) pair is called the true positive (TP), while unmatched predictions and GTs are called false positives (FP) and false negatives (FN), respectively. For IoU calculation, this invention considers t IoU The value is 0.5. The F1 score is calculated as follows:

[0083]

[0084] For the Tusimple dataset, the official evaluation metrics include accuracy, false positive rate (FPR), and false negative rate (FNR). The formula for calculating accuracy is:

[0085]

[0086] Among them, C clip S clip These are the correctly predicted pixels and the Ground Truth pixels, respectively. A predicted lane is considered correctly predicted if more than 85% of the predicted lane points are within 20 pixels of the ground truth points. The Tusimple dataset also reports the false positive rate (FPR) and false negative rate (FNR), where...

[0087] 1. Experimental Details

[0088] The experimental environment for this invention is Ubuntu 20.04.3LTS operating system, using NVIDIA GEFORCE RTX3090 graphics card with 24GB of video memory, 128GB of RAM, and an Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz. All models in this paper are based on Python 3.8 runtime environment, PyTorch 1.8.0 deep learning framework, and CUDA 11.1.

[0089] 2. Qualitative comparison

[0090] To qualitatively compare the differences between the model of this invention and the benchmark model, the lane line prediction results of the model of this invention and the benchmark model were visualized in images of the same scene, as shown in the following figures. Figure 4 As shown in the diagram. The first row represents the Ground Truth, the second row represents a visualization of the CLRNet detection results, and the third row represents a visualization of the detection results from the model of this invention. Each column represents complex scenarios such as nighttime, no lane lines, strong light, lane line occlusion and missing lines, and curves.

[0091] from Figure 4 It can be intuitively felt that, compared with the baseline model, the model of this invention has stronger performance stability in complex scenarios such as curves, strong light, lane line occlusion and missing lane lines. In such scenarios, the model of this invention can detect complete lane lines, while the baseline model can only detect half of the lane line data or even fail to detect lane lines at all.

[0092] 3. Quantitative comparison

[0093] Tables 1 and 2 show the evaluation results of the models on the TuSimple and CULane datasets, respectively. TuSimple is a small dataset with clear weather and highway scenarios and relatively easy evaluation metrics; most methods perform well on this dataset. Therefore, this invention mainly focuses on the large CULane dataset, where significant differences still exist between various methods.

[0094] Table 1. Quantitative comparison of models on the TuSimple dataset.

[0095]

[0096] Note: Bold text represents the best result, and underlined text represents the second-best result.

[0097] Table 2 Quantitative comparison of models on the CULane dataset

[0098]

[0099]

[0100] Note: Bold text represents the best result, and underlined text represents the second-best result.

[0101] As shown in Table 2, the ResNet18 version of the model of this invention achieved an F1@50 score of 79.87%, even higher than the baseline model (ResNet34). Specifically, the model of this invention achieved F1 scores more than 2% and 1% higher than the baseline model in bright light and shadow conditions, respectively. For lane-line-free scenarios, the model of this invention improved the F1 score by nearly 1%, and in curved scenarios, it improved the F1 score by nearly 3%. This indicates that the model of this invention can better adapt to complex road scenarios, especially curved ones. Compared with previous advanced lane detection models, the model of this invention has advantages in performance stability under complex scenarios.

[0102] To verify the effectiveness of the proposed module, an ablation experiment was conducted on the CULane dataset, and the results are shown in Table 3. The first row shows the experimental results of the baseline model. The second row shows the results with the Global Feature Optimizer (GFO) added to the baseline model. The last row shows the results with the Global Feature Optimizer (GFO) and the Adaptive Lane Geometry Aggregator (ALGA) added.

[0103] Table 3 shows the experimental results of model ablation on the CULane dataset.

[0104] Baseline GFO ALGA F1 √ 79.58 √ √ 79.62 √ √ √ 79.87

[0105] As can be seen from Table 3, with the addition of the Global Feature Optimizer (GFO), the model can more easily notice lane line regions in the image, thereby improving the final detection results. When both the Global Feature Optimizer (GFO) and the Adaptive Lane Geometry Aggregator (ALGA) are added, the model can not only notice key areas of the image, but also fully integrate local and global information by utilizing the special shape prior of lane lines, ultimately improving the model's detection performance.

[0106] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A lane line detection method based on global feature optimization and shape feature perception aggregation, characterized in that, Includes the following steps: Preprocess the input lane line image; Features at different levels are extracted using the backbone network; The extracted features are processed under the operation of the Global Feature Optimizer (GFO). The adaptive lane geometry aggregator ALGA utilizes prior information about the shape of lane lines to enhance the relevance of lane anchor points in the image. The extracted features are refined layer by layer from coarse to fine through the neck network to optimize the extracted features and maintain high-level semantic information throughout the process. The optimized features are processed through operations including pooling, convolution, and fully connected layers, and then fed into the regression and classification branches for feature aggregation and visualization to obtain the final lane line prediction image. The adaptive lane geometry aggregator (ALGA) is specifically implemented as follows: For a lane anchor point, a convolutional layer is first used to predict its offset from K surrounding anchor points in the same lane and the corresponding tuning factor for each offset. This tuning factor allows the model to adaptively adjust the contribution of each anchor point, as shown in the following formula: Where, p i This represents the coordinates of the i-th anchor point. Let represent the feature representation of the i-th anchor point, φ represent the offset convolutional layer used to learn and extract the predicted offset from the feature map, and ψ represent the quality adjustment factor convolutional layer used to learn and extract the quality adjustment factor for each offset from the feature map. This represents the set of predicted offsets. This represents the offset of the m-th offset. This represents the set of conditioning factors indicating the predicted offset. This represents the conditioning factor for the m-th offset; Next, deformable convolution is used to integrate the features adjacent to the anchor point by combining the predicted offset and the quality adjustment factor corresponding to each offset, so as to aggregate the context information of the i-th anchor point, as shown below: Among them, w m m = 1, ..., M are the weights of the convolution; After feature aggregation through deformable convolution, spatial attention is applied to the feature map to effectively handle cases of missing or occluded lane lines, resulting in the final lane line prediction image.

2. The lane line detection method based on global feature optimization and shape feature perception aggregation according to claim 1, characterized in that, The method for preprocessing the input lane line image is to crop the lane line image to a size of 320×800.

3. The lane line detection method based on global feature optimization and shape feature perception aggregation according to claim 1, characterized in that, The backbone network is ResNet-18.

4. The lane line detection method based on global feature optimization and shape feature perception aggregation according to claim 1, characterized in that, The neck network is a feature pyramid.

5. A lane line detection method based on global feature optimization and shape feature perception aggregation according to claim 1 or 4, characterized in that, The Global Feature Optimizer (GFO) is implemented as follows: For an input feature map of size H×W×C, where H represents the height of the input feature map, W represents the width of the input feature map, and C represents the number of channels of the input feature map, the first step is to process the efficient channel attention in ECA. This first performs global feature pooling on the input feature map, resulting in a 1×1×C output. Next, a one-dimensional convolution operation is performed on this output to capture the inter-channel dependencies. The size of the convolution kernel is adaptively adjusted according to the number of channels in the input feature map, as shown in the following formula: Where |t| odd Let represent the nearest odd number to t, and set γ and b as constants; The output is then passed through a Sigmoid activation function to obtain the final channel attention features, which are then multiplied element-wise with the original feature map. This process is represented by the following formula: F eca =F×σ(Conv 1D (AvgPool(F))) AvgPool is a global average pooling method that compresses the spatial dimension of the feature map F, retaining only channel information; Conv... 1D σ is a one-dimensional convolution used to capture the dependencies between channels; σ is the Sigmoid activation function used to generate channel attention weights. Next, the feature map with dimensions H×W×C output by ECA continues to serve as input for the spatial attention of CBAM. First, max pooling and average pooling operations are performed on the feature map along the channel dimension to obtain two H×W×1 feature maps. Then, a concat operation is performed on these two feature maps along the channel dimension. Next, a convolution operation with a kernel size of 7×7 is used to reduce the channel dimension and learn the importance of different spatial locations. Then, the sigmoid function is used to generate spatial attention features, and this feature is then multiplied element-wise with the feature map output by ECA to obtain the final feature map that fuses channel and spatial attention. This process is represented by the following formula: F spatial =σ(Conv 2D (Concat(AvgPool(F eca ),MaxPool(F eca ))))×F eca Among them, AvgPool(F eca ) and MaxPool(F eca ) represent the feature maps F after ECA adjustment. eca Perform average pooling and max pooling; Concat concatenates the results of the two pooling operations along the channel dimension; Conv 2D σ is a two-dimensional convolution operation used to learn the importance of different spatial locations; σ is the sigmoid activation function used to generate spatial attention weights.

6. The lane line detection method based on global feature optimization and shape feature perception aggregation according to claim 5, characterized in that, γ and b are set to 2 and 1, respectively.

7. The lane line detection method based on global feature optimization and shape feature perception aggregation according to claim 1, characterized in that, The Adaptive Lane Geometry Aggregator (ALGA) is used to perceive the shape features of lane lines. It predicts the offset between anchor points and their surrounding anchor points, as well as the tuning factor of each offset. It then uses deformable convolution to combine these offsets and tuning factors to integrate features and aggregate the contextual information of the anchor points.

Images