Lane detection method, device and equipment and storage medium

Abstract

The invention provides a lane detection method. Comprising the steps of obtaining laser radar point cloud data of a vehicle, performing feature extraction based on the laser radar point cloud data to obtain a feature sequence, performing feature embedding by using a Transform model based on the feature sequence, initializing a generated lane query model by using the encoded point cloud feature sequence, and performing lane query based on the initialized lane query model. Inputting the lane query features and the lane features into a multi-layer perceptron for lane detection to obtain a lane detection result, wherein the lane query features are obtained through the initialized generated lane query model; the lane features are obtained by inputting the encoded point cloud feature sequence and the lane query features into a Transform model decoder. Through the combination of the three-dimensional sparse convolutional network and the Transform model, the rapid and accurate feature extraction of the point cloud data is realized, then the lane detection is performed by using the multilayer perceptron, the information of the input features is fully utilized, and the lane detection precision is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a lane detection method, device, equipment and storage medium. Background Art

[0002] In the continuous evolution of advanced driver assistance systems (ADAS) and autonomous driving systems (ADS), lane detection technology, as a core component, plays a vital role. It is not only the basis for realizing lane departure warning, automatic lane center keeping, intelligent lane change operation, and providing effective auxiliary functions in traffic congestion, but also directly related to the overall improvement of driving safety and driving experience.

[0003] Traditional lane detection relies on manually designed rules. Faced with complex and changeable actual road conditions, such as weather changes and road conditions, it is unable to cope with the lack of flexibility and generalization ability. Deep learning technology, especially image-based lane detection, has significantly improved the ability of automatic feature learning, but is plagued by problems such as lighting conditions, two-dimensional information limitations, and image distortion. Solutions based on laser radar (LiDAR) have become a new research focus because they provide high-precision three-dimensional information and are not affected by lighting, showing higher environmental adaptability. However, current LiDAR lane detection technologies mostly rely on segmentation methods, have limited ability to handle complex scenes, and the efficiency and accuracy of existing models in processing large-scale point cloud data still need to be improved.

[0004] Therefore, how to improve the accuracy and efficiency of lane detection is a technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of this, the present application provides a lane detection method, apparatus, device and storage medium, aiming to improve the accuracy and efficiency of lane detection.

[0006] In a first aspect, the present application provides a lane detection method, the method comprising:

[0007] Get the vehicle's lidar point cloud data;

[0008] Based on the laser radar point cloud data, a three-dimensional sparse convolutional network is used to extract features to obtain a feature sequence of the laser radar point cloud data;

[0009] Based on the feature sequence of the laser radar point cloud data, an encoder of a Transformer model is used to perform feature embedding to obtain an encoded point cloud feature sequence;

[0010] Initializing a generated lane query model according to the encoded point cloud feature sequence;

[0011] The lane query feature and the lane feature are input into a multi-layer perceptron for lane detection to obtain a lane detection result; the lane query feature is obtained by generating a lane query model after initialization; the lane feature is obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder.

[0012] Optionally, the extracting features based on the laser radar point cloud data using a three-dimensional sparse convolutional network to obtain a feature sequence of the laser radar point cloud data includes:

[0013] Perform voxel feature extraction based on the laser radar point cloud data to obtain voxel feature values ​​of the laser radar point cloud data;

[0014] Inputting the voxel feature values ​​of the laser radar point cloud data into the three-dimensional sparse convolution network to perform a three-dimensional sparse convolution operation to obtain a feature map of the laser radar point cloud data;

[0015] The feature map of the laser radar point cloud data is flattened to obtain a feature sequence of the laser radar point cloud data.

[0016] Optionally, the extracting voxel features based on the laser radar point cloud data to obtain voxel feature values ​​of the laser radar point cloud data includes:

[0017] Predefine spatial characteristic attributes of unit voxels; the spatial characteristic attributes include the length, width and height corresponding to the unit voxels;

[0018] Divide the three-dimensional space where the laser radar point cloud data is located into a plurality of voxels according to the length, width and height corresponding to the unit voxel;

[0019] Allocating each point in the laser radar point cloud data to a voxel corresponding to each point according to a preset rule; the preset rule is determined by comparing the coordinates of the point cloud data within the unit voxel with the voxel boundary;

[0020] The voxel feature is calculated based on the distribution of the laser radar point cloud data in the voxel to obtain the voxel feature value.

[0021] Optionally, the output layer of the encoder of the Transformer model is connected to the input layer of the generated lane query model, and the generated lane query model is initialized using the encoded point cloud feature sequence, including:

[0022] Inputting the encoded point cloud feature sequence into the input layer of the lane query generation model through the output layer of the encoder of the Transformer model;

[0023] The weight matrix in the lane query model is initialized using the encoded point cloud feature sequence.

[0024] Optionally, the multilayer perceptron includes at least a first multilayer perceptron and a second multilayer perceptron, and the step of inputting the lane query feature and the lane feature into the multilayer perceptron to perform lane detection to obtain a lane detection result includes:

[0025] Using the first multi-layer perceptron to perform lane classification confidence detection based on the lane query feature and the lane feature to obtain a lane classification detection result;

[0026] Using the second multi-layer perceptron to perform lane position parameter detection based on the lane query feature and the lane feature to obtain a lane position parameter detection result;

[0027] A lane detection result is obtained based on the lane confidence detection result and the lane position parameter detection result.

[0028] Optionally, the first multi-layer perceptron is defined as follows:

[0029] A multilayer perceptron classifier is defined; the multilayer perceptron classifier comprises N fully connected layers, wherein each fully connected layer is followed by an activation function, the last layer of the N fully connected layers comprises a number of neurons equal to the number of lane line categories, and N is a positive integer;

[0030] Initialize the multi-layer perceptron classifier to obtain a first multi-layer perceptron.

[0031] Optionally, the second multi-layer perceptron is defined as follows:

[0032] A multilayer perceptron regressor is defined; the multilayer perceptron regressor comprises M fully connected layers, wherein each fully connected layer is followed by an activation function, the last layer of the M fully connected layers comprises a number of neurons equal to the number of parameters of the lane line equation, and M is a positive integer;

[0033] Initialize the multilayer perceptron regressor to obtain a second multilayer perceptron.

[0034] Optionally, the using the second multi-layer perceptron to perform lane position parameter detection based on the lane query feature and the lane feature to obtain a lane position parameter detection result includes:

[0035] Using the second multi-layer perceptron based on the lane query feature and the lane feature, determining the cubic polynomial parameters of at least one lane line through algorithm analysis and extracting the position information of the first and last points of the parking space curve;

[0036] A lane position parameter detection result is obtained based on the cubic polynomial parameters and the position information of the first and last points of the parking space curve.

[0037] In a third aspect, the present application provides a lane detection device, the device comprising:

[0038] An acquisition module is used to acquire the laser radar point cloud data of the vehicle;

[0039] A feature extraction module, used to extract features based on the laser radar point cloud data using a three-dimensional sparse convolutional network to obtain a feature sequence of the laser radar point cloud data;

[0040] A feature embedding module, used for performing feature embedding based on the feature sequence of the laser radar point cloud data using an encoder of a Transformer model to obtain an encoded point cloud feature sequence;

[0041] An initialization module, used to initialize the generated lane query model according to the encoded point cloud feature sequence;

[0042] The detection module is used to input the lane query feature and the lane feature into a multi-layer perceptron to perform lane detection to obtain a lane detection result; the lane query feature is obtained by generating a lane query model after initialization; the lane feature is obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder.

[0043] In a third aspect, the present application provides a device comprising a memory and a processor, wherein the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device performs the lane detection method described in any one of the first aspects above.

[0044] In a fourth aspect, the present application provides a computer storage medium having codes stored therein. When the codes are executed, a device executing the codes implements the lane detection method described in any one of the first aspects above.

[0045] The present application provides a lane detection method. When executing the method, the laser radar point cloud data of the vehicle is first obtained; then, based on the laser radar point cloud data, a three-dimensional sparse convolutional network is used to extract features to obtain a feature sequence of the laser radar point cloud data; then, based on the feature sequence of the laser radar point cloud data, a Transformer model encoder is used to embed features to obtain an encoded point cloud feature sequence; then, the encoded point cloud feature sequence is used to initialize the generated lane query model; finally, the lane query feature and the lane feature are input into a multi-layer perceptron for lane detection to obtain a lane detection result; the lane query feature is obtained by generating the generated lane query model after initialization; the lane feature is obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder. In this way, by directly outputting lane parameters from the laser radar point cloud data, there is no need for intermediate steps such as lane point recognition and fitting, which simplifies the processing flow and improves the detection efficiency. At the same time, the output of the Transformer encoder is used to initialize the generated lane query model, so that the model can accurately capture the geometric features and position information of the lane, and the multi-layer perceptron is used for lane detection, which can make full use of the information of the input features, thereby improving the accuracy of lane detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A flow chart of a lane detection method provided in an embodiment of the present application;

[0048] Figure 2 A schematic diagram of a possible lane detection system structure provided in an embodiment of the present application;

[0049] Figure 3 A flowchart of a possible implementation method of step S2 provided in an embodiment of the present application;

[0050] Figure 4 A schematic diagram of the encoder structure of a Transformer model provided in an embodiment of the present application;

[0051] Figure 5 A schematic diagram of a multi-layer perceptron provided in an embodiment of the present application;

[0052] Figure 6 A flowchart of a possible implementation method of step S5 provided in an embodiment of the present application;

[0053] Figure 7 A schematic diagram of the structure of a lane detection device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] As mentioned above, lane detection plays a vital role in the field of advanced driver assistance (ADAS) and autonomous driving (ADS), providing necessary perception input for multiple core functions. However, both traditional methods and the current mainstream deep learning (DL)-based methods have their own problems and challenges. Traditional methods in lane detection are highly dependent on artificially designed rules and have poor generalization performance. Image-based methods are easily affected by environmental factors, lack depth information, and are prone to distortion due to perspective conversion. LiDAR-based methods mainly rely on segmentation methods, and existing network models do not perform well when dealing with complex road scenes (such as blurred road boundaries, occluded lane lines, etc.).

[0055] In view of this, the present application provides a lane detection method. When executing the method, the laser radar point cloud data of the vehicle is first obtained; then, based on the laser radar point cloud data, a three-dimensional sparse convolutional network is used to perform feature extraction to obtain a feature sequence of the laser radar point cloud data; then, based on the feature sequence of the laser radar point cloud data, a Transformer model encoder is used to perform feature embedding to obtain an encoded point cloud feature sequence; then, the encoded point cloud feature sequence is used to initialize the generated lane query model; finally, the lane query feature and the lane feature are input into a multi-layer perceptron for lane detection to obtain a lane detection result; the lane query feature is obtained by the initialized generated lane query model; the lane feature is obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder.

[0056] In this way, by directly outputting lane parameters from the lidar point cloud data, there is no need for intermediate steps such as lane point identification and fitting, which greatly simplifies the processing flow and improves detection efficiency. At the same time, the output of the Transformer encoder is used to initialize the generation of the lane query model, so that the model can be optimized for specific lane tasks. This design enables the model to more accurately capture the geometric features and position information of the lane, improving the accuracy of lane detection.

[0057] It should be noted that the embodiments of the present application do not limit the execution subject of the lane detection method. For example, the lane detection method of the embodiments of the present application can be applied to data processing devices such as terminal devices or servers. The terminal device can be a smart phone, a computer, a personal digital assistant (PDA) or a tablet computer. The server can be an independent server, a cluster server or a cloud server.

[0058] In order to be able to understand the characteristics and technical contents of the embodiments of the present disclosure in more detail, the implementation of the embodiments of the present disclosure is described in detail below in conjunction with the accompanying drawings. The attached drawings are for reference only and are not used to limit the embodiments of the present disclosure. In the following technical description, for the convenience of explanation, a full understanding of the disclosed embodiments is provided through multiple details. However, one or more embodiments can still be implemented without these details. In other cases, in order to simplify the drawings, well-known structures and devices can be simplified for display. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0059] Figure 1 A flow chart of a lane detection method provided in an embodiment of the present application. Figure 1 As shown, the lane detection method provided in the embodiment of the present application may include:

[0060] S1, obtain the vehicle's lidar point cloud data.

[0061] LiDAR acquires three-dimensional information about the surrounding environment by emitting laser pulses and receiving the signals reflected back. These data are recorded in the form of points, forming a "point cloud". Each point in the point cloud contains its spatial position (x, y, z coordinates) and possible other information, such as reflection intensity.

[0062] After installing LiDAR on a vehicle, it will continuously scan the surrounding environment as the vehicle moves and generate point cloud data in real time. This data can be used for subsequent processing and analysis to support various automated tasks, such as road recognition, obstacle detection, lane keeping, etc.

[0063] In this embodiment, a laser radar sensor installed on a vehicle is used to obtain three-dimensional information about the vehicle's surroundings by emitting a laser beam and receiving the reflected signal. Specifically, the laser radar continuously emits laser pulses and receives light signals reflected from the surrounding environment. The reflected light signals are converted into electrical signals and processed to generate point cloud data. These data contain a large number of three-dimensional coordinate points about the surrounding environment.

[0064] S2, using a three-dimensional sparse convolutional network to extract features based on the lidar point cloud data to obtain a feature sequence of the lidar point cloud data.

[0065] 3D Sparse Convolutional Network is a deep learning model used to process sparsely distributed data in three-dimensional space. Compared with traditional convolutional neural networks, it can more effectively handle situations where most areas are empty (i.e., there are no data points) because it only calculates those areas that contain data. In this embodiment, since the lidar point cloud is usually sparse, that is, most of the space is empty, a sparse representation is constructed to reduce the amount of calculation and memory usage. Sparse representation can be achieved by converting the point cloud data into a sparse matrix or a sparse tensor, in which only the position and feature information of non-empty points are stored.

[0066] Feature extraction refers to selecting or constructing information from raw data that helps to accomplish a specific task. In this process, 3D sparse convolutional networks are able to identify patterns in point clouds and convert these patterns into feature vectors that can be used for subsequent tasks such as object detection, classification, segmentation, etc.

[0067] Figure 2 A possible lane detection system structure diagram provided in an embodiment of the present application, combined with Figure 2 As shown, the system includes a three-dimensional sparse convolutional network, a Transformer model (including an encoder and a decoder), a lane query module and a multi-layer perceptron. The sparsely represented point cloud data is input into the three-dimensional sparse convolutional network. The network consists of multiple sparse convolutional layers, each of which contains a sparse convolution operation and an activation function. The sparse convolution operation is optimized for sparse representation and only calculates the convolution between non-empty points and points in their neighborhood. This greatly improves the computational efficiency. Through layer-by-layer convolution and activation, the network is able to extract deep features of the point cloud. At the output of the network, an extracted feature sequence is obtained. These feature sequences usually contain feature information such as the geometry of the point cloud, surface normals, and curvature. It should be noted that the length of the feature sequence depends on the scale of the input point cloud data and the depth of the network.

[0068] In this embodiment, the three-dimensional sparse convolutional network is optimized for the sparsity of point cloud data, which significantly improves the computational efficiency of feature extraction. Compared with traditional dense convolutional networks, sparse convolutional networks can save a lot of computing resources and time. And the three-dimensional sparse convolutional network can extract high-precision feature information. These feature information are of great value for subsequent classification, recognition or tracking tasks. In addition, the three-dimensional sparse convolutional network can process large-scale point cloud data and can be extended to other types of three-dimensional data processing tasks. This makes the method have broad application prospects in the fields of autonomous driving, digital cities, etc. And because the lidar point cloud data has high accuracy and stability, the feature sequence obtained by feature extraction based on the data also has good robustness. This enables the method to maintain stable performance under different environments and conditions.

[0069] In an alternative embodiment, Figure 3 A possible implementation flow chart of step S2 provided in the embodiment of the present application. Figure 3 As shown, based on the laser radar point cloud data, a feature sequence of the laser radar point cloud data is obtained by extracting features using a three-dimensional sparse convolutional network, which may include:

[0070] S21, performing voxel feature extraction based on the laser radar point cloud data to obtain voxel feature values ​​of the laser radar point cloud data.

[0071] Voxel feature extraction is a method that divides LiDAR point cloud data into small cubic units (voxels) and extracts and analyzes the features of the point cloud data within each voxel. This method helps reduce the complexity of data processing and improves the efficiency and accuracy of feature extraction.

[0072] In this embodiment, point cloud voxelization is first performed based on the LiDAR point cloud data. Specifically, the three-dimensional space is divided into a series of small cubic units (voxels). The size of the voxel can be adjusted as needed to balance the computational efficiency and the accuracy of feature representation. Then each point in the point cloud data is assigned to the corresponding voxel according to its spatial position.

[0073] Then, voxel features are calculated based on the voxelized LiDAR point cloud data to obtain voxel feature values. Specifically, statistical analysis is performed on the point cloud data in each voxel, such as calculating statistics such as average height, standard deviation, maximum / minimum height, etc. Geometric features such as planarity and curvature are extracted based on the distribution and shape of the point cloud data in the voxel. The extracted voxel features are represented as a feature vector or feature map for subsequent processing and analysis.

[0074] In an optional embodiment, voxel feature extraction based on the lidar point cloud data to obtain voxel feature values ​​of the lidar point cloud data may include: pre-defining spatial feature attributes of unit voxels; the spatial feature attributes include length, width and height corresponding to the unit voxel; dividing the three-dimensional space where the lidar point cloud data is located into multiple voxels according to the length, width and height corresponding to the unit voxel; allocating each point in the lidar point cloud data to the voxel corresponding to each point according to a preset rule; the preset rule is determined by comparing the coordinates of the point cloud data in the unit voxel with the voxel boundary; and calculating the voxel features based on the distribution of the lidar point cloud data in the voxel to obtain the voxel feature value.

[0075] In an optional embodiment, the spatial characteristic attribute of the predefined unit voxel may also include the number of point cloud data in the unit voxel, and the preset rule may be set according to the number of point cloud data in the unit voxel.

[0076] The number of point cloud data within a unit voxel is an important feature that reflects the point cloud density within the voxel area. By including this attribute, point cloud data in three-dimensional space can be represented more efficiently, especially when the point cloud density is uneven. The preset rules can be adjusted according to the number of point cloud data within the unit voxel to control the granularity of data processing. For example, if the number of point cloud data within a voxel exceeds a certain threshold, you can choose to further subdivide the voxel or adopt other strategies to reduce data redundancy. By pre-setting a limit on the number of point cloud data within a unit voxel and taking this limit into account when assigning points, it is possible to avoid situations where some voxels are too crowded and other voxels are too sparse, which helps improve the efficiency of subsequent feature extraction.

[0077] In an optional embodiment, if the point cloud data has been semantically segmented, the semantic features of the voxels can also be extracted, such as whether the voxel belongs to a road, a building, etc.

[0078] In this embodiment, voxel feature extraction is performed based on the lidar point cloud data, the point cloud data is divided into small cubic units (voxels), and feature extraction and analysis are performed on the data within each voxel, which can effectively reduce the processing complexity of the point cloud data while improving the efficiency and accuracy of feature extraction.

[0079] S22, input the voxel feature values ​​of the lidar point cloud data into a three-dimensional sparse convolutional network to perform a three-dimensional sparse convolution operation to obtain a feature map of the lidar point cloud data.

[0080] In this embodiment, a three-dimensional sparse convolution operation is applied to the voxel eigenvalues. Since point cloud data is sparse in nature (i.e., most voxels are empty or contain very few points), traditional dense convolution will waste a lot of computing resources on empty voxels. Therefore, the use of three-dimensional sparse convolution can more efficiently process such sparse data. Specifically, the voxel eigenvalues ​​extracted in step S21 are used as input and input into the three-dimensional sparse convolution network. The three-dimensional sparse convolution network uses sparse convolution operations and only calculates the convolution between non-empty voxels and voxels in their neighborhood, thereby greatly improving the computational efficiency. Through layer-by-layer convolution and activation operations, the network can learn the deep features of point cloud data and generate feature maps. These feature maps contain the spatial structure and semantic information of point cloud data, providing strong support for subsequent tasks.

[0081] Specifically, before performing three-dimensional sparse convolution, define a suitable convolution kernel (or filter). Parameters such as the size, step size, and filling method of the convolution kernel are defined according to requirements, and this embodiment does not limit this. Then a sparse convolution operation is performed. For each non-empty voxel, the voxels in its neighborhood are found according to its spatial position (these neighborhood voxels may include empty voxels and non-empty voxels), and the convolution kernel is convolved with the eigenvalues ​​of these neighborhood voxels to obtain a new eigenvalue of the voxel. It should be noted that in sparse convolution, the eigenvalues ​​of empty voxels are usually regarded as zero or ignored. Finally, the new eigenvalues ​​obtained after sparse convolution of all voxels are merged or reorganized. For example, the new eigenvalues ​​are rearranged into a new three-dimensional feature map, in which each position corresponds to a voxel in the original voxel grid.

[0082] S23, flattening the feature map of the laser radar point cloud data to obtain a feature sequence of the laser radar point cloud data.

[0083] The feature map generated in step S22 is flattened to convert it into a one-dimensional feature sequence. The flattening operation is the process of arranging each element in the feature map into a one-dimensional array in a certain order. Through the flattening operation, the feature map can be converted into a data form that is easier to process and analyze, providing input for subsequent tasks such as classification, recognition or tracking.

[0084] In this embodiment, the sparsity of point cloud data is utilized to adopt three-dimensional sparse convolution to reduce the amount of calculation and improve processing efficiency. At the same time, more abundant feature information is extracted through convolution operation to improve the accuracy of lane detection.

[0085] S3, based on the feature sequence of the lidar point cloud data, the encoder of the Transformer model is used to perform feature embedding to obtain the encoded point cloud feature sequence.

[0086] The Transformer model is a deep learning model based on the self-attention mechanism, which performs well in processing sequence data. In this embodiment, the Transformer model is used to process the prepared voxel feature value sequence. Through its self-attention mechanism, the Transformer model can capture the dependencies between feature values ​​at different positions in the sequence and generate an embedded representation containing rich contextual information. This embedded representation is a high-dimensional vector that not only contains the information of the original voxel feature values, but also incorporates the interactions and relationships between the feature values. Therefore, the accuracy of lane detection can be improved.

[0087] Combination Figure 2 As shown, the Transformer model includes an encoder and a decoder. Figure 4 A schematic diagram of the encoder structure of a Transformer model provided in an embodiment of the present application. Figure 4 As shown in the figure, the encoder of the Transformer model includes multiple encoders, each of which includes a multi-head self-attention mechanism, a feedforward neural network, a residual connection, and a normalization layer. Each encoder layer consists of two sub-layer connection structures. The first sub-layer connection structure includes a multi-head attention sub-layer, a normalization layer, and a residual connection. The second sub-layer connection structure includes a feedforward fully connected sub-layer, a normalization layer, and a residual connection. The multi-head self-attention mechanism is used to capture the dependencies between different positions in the input sequence, the feedforward neural network is used for the nonlinear transformation applied after each attention layer, and the residual connection and layer normalization are used to help the model converge faster and improve performance.

[0088] In this step, the feature sequence is input to the encoder of the Transformer model for encoding. The encoder of the Transformer model uses the self-attention mechanism to capture the dependencies between voxel feature values ​​and generate an embedding representation. This embedding representation is a higher-dimensional feature vector that contains the global features of the point cloud data. Through this process, the original voxel feature value is converted into a more compact and expressive feature.

[0089] S4, using the encoded point cloud feature sequence to initialize the generated lane query model.

[0090] In lane detection tasks, a generative lane query model is usually used to generate a preliminary estimate of the lane. A generative lane query model is a model used to generate lane queries (which may refer to some representation or parameters of lane lines). Lane queries can be used to guide subsequent lane line detection or reconstruction tasks. This model can be a classifier, regressor, or graph neural network, etc. Its task is to receive input (such as encoded point cloud features) and output relevant information about lane lines for predicting lane selection, path planning, etc.

[0091] In this embodiment, the output result of the encoder of the Transformer model, that is, the encoded point cloud feature sequence obtained in step S3, is used to initialize the generated lane query model. In this way, the rich feature information learned by the Transformer model can be passed to the generated lane query model, thereby improving the accuracy of lane detection.

[0092] In an alternative embodiment, referring to Figure 2 As shown, the output layer of the encoder of the Transformer model is connected to the input layer of the generated lane query model, and the output result of the encoder of the Transformer model is used to initialize the generated lane query model, which may include: inputting the encoded point cloud feature sequence into the input layer of the generated lane query model through the output layer of the encoder of the Transformer model; and initializing the weight matrix in the generated lane query model using the encoded point cloud feature sequence.

[0093] Most lane detection models in related technologies are based on DETR (object detection model) detection tasks, and object queries are randomly initialized and continuously optimized, which indicates that object queries are completed based on common attributes of the search data set; therefore, in the test phase, the features of object queries will be very limited. In this embodiment, the output layer of the encoder of the Transformer model is connected to the input layer of the generated lane query model, and the sentence embedding vector generated by the Transformer encoder is used to initialize part or all of the weights of the generated lane query model.

[0094] In this embodiment, first, the encoded point cloud feature sequence passes through the output layer of the encoder of the Transformer model, and then these encoded features are input into the input layer of the generated lane query model, establishing a connection between the two models so that the representation generated by the Transformer encoder can be used as the input of the lane query model. Secondly, the encoded point cloud feature sequence (or more specifically, the information learned from these features) is used to initialize the weight matrix in the generated lane query model. This means that the initial parameters of the lane query model are set based on the point cloud features processed by the Transformer encoder, which helps the model learn features related to lane line detection faster.

[0095] The sentence embedding vector generated by the Transformer encoder is a high-dimensional, dense vector representation that captures the semantic information of the input text. When such a vector is applied to the generative lane query model, the performance of the generative lane query model can be improved. Specifically, the sentence embedding vector obtained by the Transformer is directly connected to the input layer of the generative lane query model. For example, if the generative lane query model is a neural network, then this high-dimensional vector will be used as the first input layer of the model, and subsequent layers will learn more abstract features based on this vector to make lane-related predictions or classifications. In this case, the sentence embedding vector is used to initialize the weight matrix of a layer in the model. For example, before a fully connected layer, if the layer is intended to further process text features, these embedding vectors can be used to initialize the weights of the layer, so that the model starts learning from a starting point with a good understanding of the text. Using the sentence embedding of the Transformer directly to initialize such feature embedding layers allows the model to have a better initial understanding of the input lane or path description.

[0096] In this embodiment, the pre-training capability of Transformer is utilized to allow the generative lane query model to start learning from a state that already contains rich semantic information, thereby reducing the number of iterations required for training from random initialization and potentially improving the performance and generalization capability of lane detection.

[0097] S5, inputting the lane query feature and the lane feature into a multi-layer perceptron to perform lane detection to obtain a lane detection result.

[0098] In this embodiment, the lane query feature is obtained by initializing the generated lane query model. The generated lane query model can be built based on a neural network, and its input can be encoded point cloud features or other related information, and its output is a set of lane query features. These features can be regarded as an abstract representation of the lane line, which is used to guide the subsequent lane detection process. The generated lane query model initialized in step S4 will generate a preliminary estimate of the lane. Then, based on this preliminary estimate, a series of post-processing steps (such as optimization, filtering, etc.) are performed to obtain the final lane detection result. This result contains key information such as the location and shape of the lane.

[0099] The lane features are obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder. The decoder of the Transformer model is usually used for generation tasks, which can generate a series of output features based on the input features (here, the encoded point cloud features and lane query features). These output features are called lane features, which contain more specific and detailed lane line information.

[0100] Multilayer Perceptron is a feedforward neural network consisting of multiple fully connected layers. Here, MLP is used as the final step of lane detection. Its input is lane query features and lane features, and its output is lane detection results. Lane detection results may include information such as the position, direction, and curvature of the lane line, which can be used to guide the vehicle's driving.

[0101] In an alternative embodiment, Figure 5 A schematic diagram of a multilayer perceptron provided in an embodiment of the present application, wherein the multilayer perceptron comprises at least a first multilayer perceptron and a second multilayer perceptron. Figure 6 A possible implementation flow chart of step S5 provided in the embodiment of the present application, combined with Figure 6 As shown, the lane query feature and the lane feature are input into a multi-layer perceptron to perform lane detection to obtain a lane detection result, which may include:

[0102] S51: Utilize the first multi-layer perceptron to perform lane classification confidence detection based on the lane query feature and the lane feature to obtain a lane classification detection result.

[0103] In this embodiment, after using a deep learning model to generate a lane query model to predict lane information, two specially designed "detection heads" (a first multi-layer perceptron and a second multi-layer perceptron) are further used to refine and distinguish different aspects of the prediction task.

[0104] Specifically, the lane query features and lane features are used as input and sent to the first multi-layer perceptron (MLP) for processing. The first multi-layer perceptron performs lane classification confidence detection based on these features and gives corresponding confidence. This process usually involves multiple fully connected layers, performs feature transformation through nonlinear activation functions, and finally outputs lane classification detection results. In this embodiment, the first multi-layer perceptron can use lane query features and lane features to perform lane classification and improve the accuracy of lane detection. By outputting confidence, the reliability of the lane classification results can be evaluated, providing useful information for subsequent processing.

[0105] In an optional embodiment, the first multilayer perceptron is defined as follows: a multilayer perceptron classifier is defined; the multilayer perceptron classifier includes N fully connected layers, wherein each fully connected layer is followed by an activation function, and the last layer of the N fully connected layers contains a number of neurons equal to the number of lane line categories, and N is a positive integer; the multilayer perceptron classifier is initialized to obtain the first multilayer perceptron.

[0106] In this embodiment, in the lane line classification task, a class named MLPClassifier is defined, which is specifically used to process the decoded feature vector and output the classification result of the lane line. The specific definition is as follows: Class name: MLPClassifier. Composition: Contains three fully connected layers (fc1, fc2, fc3) and ReLU activation function. Each fully connected layer is followed by a ReLU activation function to increase the nonlinear expression ability of the model. Output layer: The output size of the last layer is equal to the number of lane line categories. For example, in a two-classification problem, the output layer can contain 2 neurons, representing the confidence of the two categories respectively.

[0107] Then initialize the MLP classifier. When initializing the MLPClassifier class, you need to specify the following parameters: input_dim: The dimension of the decoded feature vector. This is the number of neurons in the input layer, which is used to receive the feature vector output by the decoder. hidden_dim: The number of neurons in the hidden layer. This is an adjustable parameter that can be adjusted according to the specific task and the size of the dataset to optimize the performance of the model. num_classes: The number of lane line categories. This is the number of neurons in the output layer, which is used to represent different categories of lane lines.

[0108] The initialization process can specifically include: creating an input layer with the number of neurons equal to input_dim. Creating three fully connected layers (fc1, fc2, fc3), each with the number of neurons hidden_dim (for fc1 and fc2) and numclasses (for fc3). Adding a ReLU activation function after each fully connected layer. Connecting these layers in order to form a complete MLP network.

[0109] S52: Utilize the second multi-layer perceptron to perform lane position parameter detection based on the lane query feature and the lane feature to obtain a lane position parameter detection result.

[0110] In this embodiment, in parallel or after the confidence detection, the second multi-layer perceptron works to accurately locate the position of the lane line and extract its geometric parameters, such as the centerline position, slope, width, etc. of the lane line. This step not only focuses on whether there is a lane, but also on the specific shape of the lane line and its precise position in the image. The output usually contains a series of parameters or bounding boxes that describe the specific position and shape of each detected lane line.

[0111] In an optional embodiment, using the second multi-layer perceptron to perform lane position parameter detection based on the lane query feature and the lane feature to obtain a lane position parameter detection result may include: using the second multi-layer perceptron to determine the cubic polynomial parameters of at least one lane line based on the lane query feature and the lane feature through algorithm analysis and extracting the position information of the first and last points of the parking space curve; obtaining the lane position parameter detection result based on the cubic polynomial parameters and the position information of the first and last points of the parking space curve.

[0112] In this embodiment, the task of the second multi-layer perceptron is to deeply analyze the feature information provided by the generated lane query model to identify and quantify the specific location of the lane line. This usually involves identifying the center line or edge of the lane line, measuring the width of the lane, calculating the inclination angle of the lane line, etc. The detection result may include a series of parameters, such as the slope and intercept of the straight line equation, or the fitting parameters of the curve.

[0113] Specifically, when the lane shape is complex, especially in the case of curved lanes, the cubic polynomial (generally in the form of y = ax 3 +bx 2 +cx+d) is used to accurately describe the curve shape of the lane line. This polynomial can adapt well to lane lines of various curvatures, and fit the curvature changes of the lane line by adjusting the parameters a, b, c and d. These parameters are calculated by the algorithm to ensure a high degree of fit to the shape of the lane line.

[0114] In addition to fitting the overall shape of the lane line, the precise start and end positions are crucial to understanding the complete layout of the lane line. These position information ensures that the fitted curve not only has the right shape, but also correctly locates the start and end of the lane line in space, which is particularly critical for functions such as vehicle path planning and parking assistance. Therefore, in this embodiment, the precise position information of the first and last points of the parking space curve extracted from the image data is combined to ensure that the start and end points of the lane line match the actual lane.

[0115] In this embodiment, the second multi-layer perceptron not only determines the basic position of the lane line based on the output of the generated lane query model, but also accurately depicts the curve shape of the lane line through cubic polynomial parameters, and combines the first and last point position information to ensure the integrity and accuracy of the lane line. Such detection results provide high-precision input data for subsequent lane tracking and path planning algorithms, improving the reliability and efficiency of lane recognition in autonomous driving or intelligent transportation systems.

[0116] In an optional embodiment, the second multilayer perceptron is defined as follows: a multilayer perceptron regressor is defined; the multilayer perceptron regressor includes M fully connected layers, wherein each fully connected layer is followed by an activation function, and the last layer of the M fully connected layers contains a number of neurons equal to the number of lane line equation parameters, and M is a positive integer; and the multilayer perceptron regressor is initialized to obtain the second multilayer perceptron.

[0117] Similar to the definition process of the second multi-layer perceptron, in this embodiment, in the lane line equation parameter estimation task, a class named MLPRegressor is defined, which is specifically used to process the decoded feature vector and output the regression parameters of the lane line equation. The specific definition is as follows: Class name: MLPRegressor. Composition: Contains three fully connected layers (fc1, fc2, fc3) and a ReLU activation function. Each fully connected layer is followed by a ReLU activation function to increase the nonlinear expression ability of the model. Output layer: The output size of the last layer is equal to the number of parameters of the lane line equation. For example, if the lane line equation is a quadratic curve, then three parameters (a, b, c) may need to be output to represent the curve.

[0118] Then initialize the MLP regressor. When initializing the MLPRegressor class, you need to specify the following parameters: input_dim: The dimension of the decoded feature vector. This is the number of neurons in the input layer, which is used to receive the feature vector output by the decoder. hidden_dim: The number of neurons in the hidden layer. This is an adjustable parameter that can be adjusted according to the specific task and the size of the dataset to optimize the performance of the model. output_dim: The number of parameters of the lane equation. This is the number of neurons in the output layer, which is used to represent the parameters of the lane equation.

[0119] The initialization process specifically includes: creating an input layer with the number of neurons equal to input_dim. Creating three fully connected layers (fc1, fc2, fc3), each with the number of neurons hidden_dim (for fc1 and fc2) and output_dim (for fc3). Adding a ReLU activation function after each fully connected layer. Connecting these layers in order to form a complete MLP network.

[0120] S53: Obtain a lane detection result based on the lane confidence detection result and the lane position parameter detection result.

[0121] In this embodiment, the lane classification detection result output by the first multi-layer perceptron and the lane position parameter detection result output by the second multi-layer perceptron are fused. The final lane detection result is obtained by comprehensively considering the lane classification confidence and the lane position parameters. This process may include post-processing of the detection results, such as smoothing filtering, morphological operations, etc., to further optimize the shape and continuity of the lane line. That is, combining the lane classification results and the exact position parameters to determine which lane line predictions are reliable and accurately mark their positions on the image. In an optional embodiment, it may also include filtering out low-confidence detection results, solving problems such as overlap or conflict between lane lines, and thus providing clear and accurate lane layout information.

[0122] In the embodiment of the present application, lane parameters are directly output from the lidar point cloud data without intermediate steps such as lane point identification and fitting, which simplifies the processing flow and improves detection efficiency. At the same time, it removes the errors that may be introduced by the intermediate steps and improves the accuracy and robustness of lane detection. The output of the Transformer encoder is used to initialize the generation of a lane query model, so that the model can accurately capture the geometric features and position information of the lane, and the multi-layer perceptron is used for lane detection, which can make full use of the information of the input features, thereby improving the accuracy of lane detection.

[0123] The above are some specific implementations of the lane detection method provided in the embodiment of the present application. Based on this, the present application also provides a corresponding device. The device provided in the embodiment of the present application will be introduced from the perspective of functional modularization.

[0124] See also Figure 7 The schematic diagram of the structure of the lane detection device 700 shown in the figure includes an acquisition module 710, a feature extraction module 720, a feature embedding module 730, an initialization module 740 and a detection module 750.

[0125] An acquisition module 710 is used to acquire laser radar point cloud data of the vehicle;

[0126] A feature extraction module 720 is used to extract features based on the laser radar point cloud data using a three-dimensional sparse convolutional network to obtain a feature sequence of the laser radar point cloud data;

[0127] A feature embedding module 730 is used to perform feature embedding based on the feature sequence of the laser radar point cloud data using an encoder of a Transformer model to obtain an encoded point cloud feature sequence;

[0128] An initialization module 740, configured to initialize a generated lane query model according to the encoded point cloud feature sequence;

[0129] The detection module 750 is used to input the lane query feature and the lane feature into a multi-layer perceptron for lane detection to obtain a lane detection result; the lane query feature is obtained by generating a lane query model after initialization; the lane feature is obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder.

[0130] Optionally, the feature extraction module includes:

[0131] A voxel feature extraction unit, used for performing voxel feature extraction based on the laser radar point cloud data to obtain voxel feature values ​​of the laser radar point cloud data;

[0132] A convolution unit, used for inputting the voxel feature values ​​of the laser radar point cloud data into the three-dimensional sparse convolution network to perform a three-dimensional sparse convolution operation to obtain a feature map of the laser radar point cloud data;

[0133] The feature map flattening unit is used to flatten the feature map of the laser radar point cloud data to obtain a feature sequence of the laser radar point cloud data.

[0134] Optionally, the voxel feature extraction unit is specifically used to:

[0135] Predefine spatial characteristic attributes of unit voxels; the spatial characteristic attributes include the length, width and height corresponding to the unit voxels;

[0136] Divide the three-dimensional space where the laser radar point cloud data is located into a plurality of voxels according to the length, width and height corresponding to the unit voxel;

[0137] Allocating each point in the laser radar point cloud data to a voxel corresponding to each point according to a preset rule; the preset rule is determined by comparing the coordinates of the point cloud data within the unit voxel with the voxel boundary;

[0138] The voxel feature is calculated based on the distribution of the laser radar point cloud data in the voxel to obtain the voxel feature value.

[0139] Optionally, the output layer of the encoder of the Transformer model is connected to the input layer of the lane query generation model, and the initialization module is specifically used to:

[0140] Inputting the encoded point cloud feature sequence into the input layer of the lane query generation model through the output layer of the encoder of the Transformer model;

[0141] The weight matrix in the lane query model is initialized using the encoded point cloud feature sequence.

[0142] Optionally, the multilayer perceptron includes at least a first multilayer perceptron and a second multilayer perceptron, and the detection module may include:

[0143] A lane classification detection unit, configured to use the first multi-layer perceptron to perform lane classification confidence detection based on the lane query feature and the lane feature to obtain a lane classification detection result;

[0144] a lane position parameter detection unit, configured to use the second multi-layer perceptron to perform lane position parameter detection based on the lane query feature and the lane feature to obtain a lane position parameter detection result;

[0145] A detection unit is used to obtain a lane detection result based on the lane confidence detection result and the lane position parameter detection result.

[0146] Optionally, the first multi-layer perceptron is defined as follows:

[0147] A multilayer perceptron classifier is defined; the multilayer perceptron classifier comprises N fully connected layers, wherein each fully connected layer is followed by an activation function, the last layer of the N fully connected layers comprises a number of neurons equal to the number of lane line categories, and N is a positive integer;

[0148] Initialize the multi-layer perceptron classifier to obtain a first multi-layer perceptron.

[0149] Optionally, the second multi-layer perceptron is defined as follows:

[0150] A multilayer perceptron regressor is defined; the multilayer perceptron regressor comprises M fully connected layers, wherein each fully connected layer is followed by an activation function, the last layer of the M fully connected layers comprises a number of neurons equal to the number of parameters of the lane line equation, and M is a positive integer;

[0151] Initialize the multilayer perceptron regressor to obtain a second multilayer perceptron.

[0152] Optionally, the lane position parameter detection unit is specifically used to:

[0153] Using the second multi-layer perceptron based on the lane query feature and the lane feature, determining the cubic polynomial parameters of at least one lane line through algorithm analysis and extracting the position information of the first and last points of the parking space curve;

[0154] A lane position parameter detection result is obtained based on the cubic polynomial parameters and the position information of the first and last points of the parking space curve.

[0155] The embodiments of the present application also provide corresponding electronic devices and computer storage media for implementing the solutions provided by the embodiments of the present application.

[0156] The electronic device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the lane detection method described in any embodiment of the present application.

[0157] The computer storage medium stores codes, and when the codes are executed, a device executing the codes implements the lane detection method described in any embodiment of the present application.

[0158] The "first" and "second" in the names such as "first" and "second" (if any) mentioned in the embodiments of the present application are only used as name identifiers and do not represent the first or second in order.

[0159] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.

[0160] Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0161] The above description is merely an exemplary embodiment of the present application and is not intended to limit the protection scope of the present application.

Claims

1. A lane detection method, characterized in that: The method comprises: Get the vehicle's lidar point cloud data; Based on the laser radar point cloud data, a three-dimensional sparse convolutional network is used to extract features to obtain a feature sequence of the laser radar point cloud data; Based on the feature sequence of the laser radar point cloud data, an encoder of a Transformer model is used to perform feature embedding to obtain an encoded point cloud feature sequence; Initializing a generated lane query model according to the encoded point cloud feature sequence; The lane query feature and the lane feature are input into a multi-layer perceptron for lane detection to obtain a lane detection result; the lane query feature is obtained by generating a lane query model after initialization; the lane feature is obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder.

2. The method according to claim 1, characterized in that The method of extracting features based on the laser radar point cloud data using a three-dimensional sparse convolutional network to obtain a feature sequence of the laser radar point cloud data includes: Perform voxel feature extraction based on the laser radar point cloud data to obtain voxel feature values ​​of the laser radar point cloud data; Inputting the voxel feature values ​​of the laser radar point cloud data into the three-dimensional sparse convolution network to perform a three-dimensional sparse convolution operation to obtain a feature map of the laser radar point cloud data; The feature map of the laser radar point cloud data is flattened to obtain a feature sequence of the laser radar point cloud data.

3. The lane detection method according to claim 2, characterized in that: The voxel feature extraction based on the laser radar point cloud data to obtain the voxel feature value of the laser radar point cloud data includes: Predefine spatial characteristic attributes of unit voxels; the spatial characteristic attributes include the length, width and height corresponding to the unit voxels; Divide the three-dimensional space where the laser radar point cloud data is located into a plurality of voxels according to the length, width and height corresponding to the unit voxel; Allocating each point in the laser radar point cloud data to a voxel corresponding to each point according to a preset rule; the preset rule is determined by comparing the coordinates of the point cloud data within the unit voxel with the voxel boundary; The voxel feature is calculated based on the distribution of the laser radar point cloud data in the voxel to obtain the voxel feature value.

4. The lane detection method according to claim 1, characterized in that: The output layer of the encoder of the Transformer model is connected to the input layer of the generated lane query model, and the generated lane query model is initialized using the encoded point cloud feature sequence, including: Inputting the encoded point cloud feature sequence into the input layer of the lane query generation model through the output layer of the encoder of the Transformer model; The weight matrix in the lane query model is initialized using the encoded point cloud feature sequence.

5. The lane detection method according to claim 1, characterized in that: The multi-layer perceptron includes at least a first multi-layer perceptron and a second multi-layer perceptron, and the step of inputting the lane query feature and the lane feature into the multi-layer perceptron to perform lane detection to obtain a lane detection result includes: Using the first multi-layer perceptron to perform lane classification confidence detection based on the lane query feature and the lane feature to obtain a lane classification detection result; Using the second multi-layer perceptron to perform lane position parameter detection based on the lane query feature and the lane feature to obtain a lane position parameter detection result; A lane detection result is obtained based on the lane confidence detection result and the lane position parameter detection result.

6. The lane detection method according to claim 5, characterized in that: The first multi-layer perceptron is defined as follows: A multilayer perceptron classifier is defined; the multilayer perceptron classifier comprises N fully connected layers, wherein each fully connected layer is followed by an activation function, the last layer of the N fully connected layers comprises a number of neurons equal to the number of lane line categories, and N is a positive integer; Initialize the multi-layer perceptron classifier to obtain a first multi-layer perceptron.

7. The lane detection method according to claim 5, characterized in that: The second multilayer perceptron is defined as follows: A multilayer perceptron regressor is defined; the multilayer perceptron regressor comprises M fully connected layers, wherein each fully connected layer is followed by an activation function, the last layer of the M fully connected layers comprises a number of neurons equal to the number of parameters of the lane line equation, and M is a positive integer; Initialize the multilayer perceptron regressor to obtain a second multilayer perceptron.

8. The lane detection method according to claim 5, characterized in that: The using the second multi-layer perceptron to perform lane position parameter detection based on the lane query feature and the lane feature to obtain a lane position parameter detection result includes: Using the second multi-layer perceptron based on the lane query feature and the lane feature, determining the cubic polynomial parameters of at least one lane line through algorithm analysis and extracting the position information of the first and last points of the parking space curve; A lane position parameter detection result is obtained based on the cubic polynomial parameters and the position information of the first and last points of the parking space curve.

9. A lane detection device, characterized in that: The device comprises: An acquisition module is used to acquire the laser radar point cloud data of the vehicle; A feature extraction module, used to extract features based on the laser radar point cloud data using a three-dimensional sparse convolutional network to obtain a feature sequence of the laser radar point cloud data; A feature embedding module, used for performing feature embedding based on the feature sequence of the laser radar point cloud data using an encoder of a Transformer model to obtain an encoded point cloud feature sequence; An initialization module, used to initialize the generated lane query model according to the encoded point cloud feature sequence; The detection module is used to input the lane query feature and the lane feature into a multi-layer perceptron to perform lane detection to obtain a lane detection result; the lane query feature is obtained by generating a lane query model after initialization; the lane feature is obtained by inputting the encoded point cloud feature sequence and the lane query feature into the Transformer model decoder.

10. An electronic device, comprising a memory and a processor, wherein the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes, so that the device executes the lane detection method according to any one of claims 1 to 8.

11. A computer storage medium, wherein a code is stored in the computer storage medium. When the code is executed, a device executing the code implements the lane detection method according to any one of claims 1 to 8.

Images