Drivable area detection method, domain controller and computer-readable storage medium

Abstract

The present invention provides a drivable area detection method, a domain controller and a computer-readable storage medium. The method includes: obtaining a road surface image; inputting the road surface image into a drivable area detection model, performing road surface segmentation on the road surface image to identify the drivable area in the road surface image; the drivable area detection model includes an encoder, a decoder and an attention module, the attention module extracts attention features from the coding features output by the encoder, and inputs the attention features into the decoder for feature fusion; the encoder adopts a residual network structure, including N sequentially connected downsampling layers; the decoder includes N sequentially connected multi-scale feature fusion layers, which are connected to the downsampling layers in a one-to-one correspondence, and each layer fuses the coding features of the corresponding downsampling layer, the fusion features of the previous layer and the attention features; the present invention fully considers global features and local features, effectively improves the detection accuracy of the drivable area, improves the inference speed of the model, and achieves the effect of real-time detection.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a drivable area detection method, a domain controller, and a computer-readable storage medium. Background Art

[0002] The automatic driving of electric vehicles is mainly achieved through the road perception system, which is mainly divided into target detection, lane line detection, drivable area segmentation, etc. Among them, the drivable area segmentation is based on the recognition of high-level semantic information at the image pixel level. The high-level semantic information generally includes drivable area boundary information, lane information, target detection information and target tracking information. Lane information, target detection information and target tracking information need to be based on drivable area boundary information. Therefore, the detection of drivable area information occupies a primary position in the road perception system. The current main drivable area detection mainly uses the Transformer model, which is mainly aimed at the extraction of global features. The extraction of local features is still lacking, resulting in low accuracy in drivable area detection. At the same time, the Transformer model has a complex structure and requires high hardware computing power, making it difficult to achieve real-time detection, making it difficult to implement and apply the algorithm. Summary of the Invention

[0003] In response to the above problems, the purpose of the present invention is to provide a drivable area detection method, a domain controller and a computer-readable storage medium, which fully consider global and local features, effectively improve the detection accuracy of the drivable area, and at the same time improve the inference speed of the model, thereby achieving real-time detection.

[0004] In a first aspect, an embodiment of the present invention provides a method for detecting a drivable area, comprising:

[0005] Acquire road surface images;

[0006] Inputting the road surface image into a drivable area detection model, performing road surface segmentation on the road surface image to identify the drivable area in the road surface image;

[0007] The drivable area detection model includes an encoder, a decoder, and an attention module. The attention module extracts attention features from the encoding features output by the encoder and inputs them into the decoder for feature fusion.

[0008] Among them, the encoder adopts a residual network structure, including N sequentially connected downsampling layers; the decoder includes N sequentially connected multi-scale feature fusion layers, and is connected one-to-one with the downsampling layers. Each layer of multi-scale feature fusion layer fuses the encoding features of the corresponding downsampling layer in the encoder, the fusion features of the previous multi-scale feature fusion layer, and the attention features output by the attention module.

[0009] As an improvement to the above solution, the edge enhancement module is provided between part of the downsampling layer and the multi-scale feature fusion layer to extract edge features.

[0010] As an improvement of the above solution, the residual network structure adopted by the encoder is ResNet34; wherein, the number of layers of the downsampling layer and the multi-scale feature fusion layer are both 4.

[0011] As an improvement to the above scheme, the input of the multi-scale feature fusion layer includes: the concatenation of the encoding features output by the corresponding downsampling layer and the edge features output by its corresponding edge enhancement module, the fusion features output by the previous multi-scale feature fusion layer, and the attention features output by the attention module.

[0012] As an improvement to the above solution, the edge enhancement module uses the HPF operator to extract edge features.

[0013] As an improvement to the above solution, the template of the HPF operator is:

[0014] As an improvement to the above solution, the function of the multi-scale feature fusion layer is expressed as follows:

[0015]

[0016] Among them, α4+α5=1, α1+α2+α3=1, f up Represents upsampling operation, MSFF i Represents the features output by the i-th multi-scale feature fusion layer, MSFF i-1 represents the features output by the i-1th multi-scale feature fusion layer, f conv1×1 represents a convolution layer with a convolution kernel size of 1, GLIBN is the attention feature output by the attention module, It is the encoded feature output by the 4-i-1th downsampling layer.

[0017] As an improvement of the above scheme, the attention module includes a first stage, a second stage, a convolutional neural network layer, a multi-head self-attention mechanism layer and a multi-layer perceptron; wherein, the input of the first stage is the encoded features output by the Nth downsampling layer; the input of the convolutional neural network layer and the multi-head self-attention mechanism layer is the features output by the first stage; the input of the second stage is the concatenation operation result of the encoded features output by the Nth downsampling layer and the features output by the convolutional neural network layer and the multi-head self-attention mechanism layer; the input of the multi-layer perceptron is the features output by the second stage; the output of the attention module is the concatenation operation result of the features input to the second stage and the features output by the multi-layer perceptron.

[0018] As an improvement to the above scheme, the first stage includes a first feature slicing layer, a first batch of normalization layers, a first instance normalization layer and a first feature splicing layer; wherein, the input of the first feature slicing layer is the encoded features output by the Nth downsampling layer, the input of the first batch of normalization layers and the first instance normalization layer is the features output by the first feature slicing layer, the input of the first feature splicing layer is the features output by the first batch of normalization layers and the first instance normalization layer, and the output of the first feature splicing layer serves as the output of the first stage.

[0019] As an improvement to the above scheme, the second stage includes a second feature slice layer, a second batch normalization layer, a second instance normalization layer and a second feature splicing layer; wherein, the input of the second feature slice layer is the input of the second stage, the input of the second batch normalization layer and the second instance normalization layer is the features output by the second feature slice layer, the input of the second feature splicing layer is the features output by the second batch normalization layer and the second instance normalization layer, and the output of the second feature splicing layer is the output of the second stage.

[0020] In a second aspect, an embodiment of the present invention provides a domain controller, including:

[0021] One or more processors for deploying a drivable area detection model;

[0022] a memory for storing one or more programs,

[0023] When the one or more programs are executed by the one or more processors, the one or more processors implement the drivable area detection method as described in any one of the first aspects.

[0024] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the drivable area detection method as described in any one of the first aspects.

[0025] Compared with the prior art, the beneficial effects of the embodiments of the present invention are: obtaining a road surface image; inputting the road surface image into a drivable area detection model, performing road segmentation on the road surface image to identify the drivable area in the road surface image; the drivable area detection model includes an encoder, a decoder and an attention module, the attention module extracts attention features from the coding features output by the encoder, and inputs the attention features into the decoder for feature fusion; wherein the encoder adopts a residual network structure, including N sequentially connected downsampling layers; the decoder includes N sequentially connected multi-scale feature fusion layers, and is connected one-to-one with the downsampling layers, each layer of multi-scale feature fusion layer fuses the coding features of the corresponding downsampling layer in the encoder, the fusion features of the previous layer of multi-scale feature fusion layer and the attention features output by the attention module. The present invention fully considers global features and local features, effectively improves the detection accuracy of the drivable area, and at the same time improves the inference speed of the model, thereby achieving the effect of real-time detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the implementation. Obviously, the drawings described below are only some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 This is a flow chart of a method for detecting a drivable area provided by an embodiment of the present invention;

[0028] Figure 2 Schematic diagram of the structure of the drivable area detection model provided by an embodiment of the present invention;

[0029] Figure 3 is a schematic diagram of the structure of the attention module provided by an embodiment of the present invention;

[0030] Figure 4 This is a schematic block diagram of the structure of a domain controller provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0032] Example 1

[0033] See also Figure 1 , an embodiment of the present invention provides a drivable area detection method, which can be executed by a road surface perception system installed on a vehicle domain controller, specifically comprising:

[0034] S1: Acquire road surface image;

[0035] When the vehicle is in autonomous driving mode or other driving scenarios, the onboard camera captures the road ahead in real time, and the road perception system acquires the road image captured by the onboard camera. Furthermore, the road image can be pre-processed, for example, by binarizing or gray-scaling it to reduce the amount of data required and improve detection efficiency. Furthermore, the road image can be scaled or cropped to obtain a fixed-size image, for example, by uniformly scaling the road image to a 512x512x3 image.

[0036] S2: Inputting the road surface image into a drivable area detection model, performing road segmentation on the road surface image to identify the drivable area in the road surface image;

[0037] The drivable area detection model includes an encoder, a decoder, and an attention module. The attention module extracts attention features from the encoding features output by the encoder and inputs them into the decoder for feature fusion.

[0038] Among them, the encoder adopts a residual network structure, including N sequentially connected downsampling layers; the decoder includes N sequentially connected multi-scale feature fusion layers, and is connected one-to-one with the downsampling layers. Each layer of multi-scale feature fusion layer fuses the encoding features of the corresponding downsampling layer in the encoder, the fusion features of the previous multi-scale feature fusion layer, and the attention features output by the attention module.

[0039] In an embodiment of the present invention, the road surface image is input into the drivable area detection model, wherein the encoding module performs feature extraction on the road surface image to obtain encoding features; the encoding features are used as inputs to the attention module and the decoding module, and the attention module focuses on local features such as detail information and style domain information of the road surface image to obtain attention features; finally, based on the encoding features and the attention features, feature fusion is performed through the decoding module to obtain the final road surface segmentation result. Combined with the attention features, the completeness of road segmentation can be enhanced, and global information and local information can be paid attention to at the same time, thereby improving the detection accuracy of the drivable area, improving the inference speed, and improving the lightweight effect of the model, which is conducive to improving the accuracy and performance of autonomous driving.

[0040] In an optional embodiment, the edge enhancement module is provided between part of the downsampling layer and the multi-scale feature fusion layer to extract edge features.

[0041] In an embodiment of the present invention, for the downsampling layer located in the middle of the encoder, its output is used as the input of the corresponding edge enhancement module, and then a splicing operation is performed on the output of the downsampling layer and the output of its corresponding edge enhancement module, and the splicing operation result is input into the corresponding multi-scale feature fusion layer, and feature fusion is performed in combination with the attention feature and the output of the previous multi-scale feature fusion layer.

[0042] In an optional embodiment, the residual network structure adopted by the encoder is ResNet34; wherein the number of layers of the downsampling layer and the multi-scale feature fusion layer are both 4.

[0043] In an optional embodiment, the input of the multi-scale feature fusion layer includes: the concatenation of the encoding features output by the corresponding downsampling layer and the edge features output by its corresponding edge enhancement module, the fusion features output by the previous multi-scale feature fusion layer, and the attention features output by the attention module.

[0044] like Figure 2As shown, a drivable area detection model consisting of 4 downsampling layers and 4 multi-scale feature fusion layers is taken as an example for explanation: the encoding module includes a ResNet34 residual network structure, including Stage-(1-4) connected in sequence, with the input of Stage-1 as the input of the encoding module and the output of Stage-4 as the output of the encoding module. Each stage has a downsampling layer for feature extraction from the input image. Each downsampling layer uses the features output by the corresponding previous downsampling layer as input. As the stage increases, the number of channels of the encoded features increases, the scale of the encoded features decreases, and the abstractness of the features increases. For example, if the road image input is a 512x512x3 image, after Stage-1, the features obtained are downsampled by 2 and have 64 channels, that is, 256x256x64. Similarly, each stage produces a downsampling, resulting in Channelx2 features. Therefore, after four layers of downsampling layers, the encoded features have the following shapes: 256x256x64, 128x128x128, 64x64x256, and 32x32x512. Using the improved residual network structure ResNet34 can achieve a more lightweight effect.

[0045] The decoding module adopts a CNN network structure, including multi-scale feature fusion layers MSFF-(1-4) connected in sequence, and the corresponding multi-scale fusion feature layers adopt upsampling operations. The input of MSFF-1 is used as the input of the decoding module, and the output of MSFF-4 is used as the output of the decoding module. Among them, the correspondence between the downsampling layer and the multi-scale feature fusion layer is: downsampling layer i corresponds to multi-scale feature fusion layer 4-i+1, that is, Stage-1 is connected to MSFF-4, Stage-2 is connected to MSFF-3 through the edge enhancement module, Stage-3 is connected to MSFF-2 through the edge enhancement module, Stage-4 is connected to MSFF-1, and the features obtained by the attention module after the output features of Stage-4 are connected to MSFF-(1-4) respectively. At this time, MSFF-1 uses the attention features input by the attention module and the coding features output by Stage-4 as input to perform the first feature fusion. MSFF- 2. A second feature fusion is performed using the fused features output by MSFF-1, the edge features output by the corresponding edge enhancement module, and the attention features input by the attention module as input. MSFF-3 uses the fused features output by MSFF-2, the edge features output by the corresponding edge enhancement module, and the attention features input by the attention module as input for a third feature fusion. MSFF-4 uses the fused features output by MSFF-3, the attention features input by the attention module, and the encoded features output by Stage-1 as input for a final feature fusion, resulting in a road segmentation result that identifies the drivable area within the road image. Using MSFF to fuse features directly mapped from the encoding module, the output features of the previous MSFF, and the attention features output by the attention module enriches the features output by the decoding module and improves its robustness. This allows for the extraction of both global and local features of the road image, making the model more adaptable and robust.

[0046] Because the Stage-1 downsampling factor is small, the image still contains noise. Edge enhancement of the Stage-1 output would increase the noise impact. Meanwhile, the Stage-4 downsampling factor is large, which causes detailed edge features to disappear. Therefore, edge enhancement of the Stage-4 output features is meaningless. Therefore, in this embodiment, an edge enhancement module is added only to the middle downsampling layer of the encoder module. This ensures that the edge features of the original road image are extracted while reducing the noise impact.

[0047] In an optional embodiment, the edge enhancement module uses a HPF operator to extract edge features.

[0048] In an optional embodiment, the template of the HPF operator is:

[0049] In an optional embodiment, the function of the multi-scale feature fusion layer is expressed as follows:

[0050]

[0051] Among them, α4+α5=1, α1+α2+α3=1, f up Represents upsampling operation, MSFF i Represents the features output by the i-th multi-scale feature fusion layer, MSFF i-1 represents the features output by the i-1th multi-scale feature fusion layer, f conv1×1 represents a convolution layer with a convolution kernel size of 1, GLIBN is the attention feature output by the attention module, It is the encoded feature output by the 4-i-1th downsampling layer.

[0052] In an optional embodiment, the attention module includes a first stage, a second stage, a convolutional neural network layer, a multi-head self-attention mechanism layer and a multi-layer perceptron; wherein, the input of the first stage is the encoded features output by the Nth downsampling layer; the input of the convolutional neural network layer and the multi-head self-attention mechanism layer is the features output by the first stage; the input of the second stage is the concatenation operation result of the encoded features output by the Nth downsampling layer and the features output by the convolutional neural network layer and the multi-head self-attention mechanism layer; the input of the multi-layer perceptron is the features output by the second stage; and the output of the attention module is the concatenation operation result of the features input to the second stage and the features output by the multi-layer perceptron.

[0053] In an optional embodiment, the first stage includes a first feature slicing layer, a first batch of normalization layers, a first instance normalization layer, and a first feature splicing layer; wherein, the input of the first feature slicing layer is the encoded features output by the Nth downsampling layer, the input of the first batch of normalization layers and the first instance normalization layer is the features output by the first feature slicing layer, the input of the first feature splicing layer is the features output by the first batch of normalization layers and the first instance normalization layer, and the output of the first feature splicing layer serves as the output of the first stage.

[0054] In an optional embodiment, the second stage includes a second feature slice layer, a second batch normalization layer, a second instance normalization layer and a second feature splicing layer; wherein the input of the second feature slice layer is the input of the second stage, the input of the second batch normalization layer and the second instance normalization layer is the features output by the second feature slice layer, the input of the second feature splicing layer is the features output by the second batch normalization layer and the second instance normalization layer, and the output of the second feature splicing layer is the output of the second stage.

[0055] like Figure 3 As shown in Figure 1, the network structure of the attention module is given, which specifically includes:

[0056] The first and second feature split layers, Split, are used to split features;

[0057] The first and second batch normalization layers, Batch Norm, are used to standardize batch features;

[0058] The first and second instance normalization layers, Instance Norm, are used to normalize single-channel features;

[0059] The first and second feature concatenation layers, Concat, are used to concatenate the features output by the batch normalization layer and the instance normalization layer;

[0060] Convolutional neural network layer Convolution Layer, used for extracting local features;

[0061] Multi-Head Self-Attention mechanism layer, used for global attention and global feature extraction;

[0062] Multi-layer perceptron MLP, used for regression or feature fitting.

[0063] Compared with the existing Transformer model that mainly extracts global features, the embodiment of the present invention can simultaneously extract global features and local features through the above-mentioned attention module. At the same time, due to the fusion of the CNN and attention mechanism of the decoding module during reasoning, the reasoning speed can be further accelerated, thereby improving the detection efficiency of the drivable area.

[0064] Compared with the prior art, the beneficial effects of the embodiments of the present invention are: obtaining a road surface image; inputting the road surface image into a drivable area detection model, performing road segmentation on the road surface image to identify the drivable area in the road surface image; the drivable area detection model includes an encoder, a decoder and an attention module, the attention module extracts attention features from the coding features output by the encoder, and inputs the attention features into the decoder for feature fusion; wherein the encoder adopts a residual network structure, including N sequentially connected downsampling layers; the decoder includes N sequentially connected multi-scale feature fusion layers, and is connected one-to-one with the downsampling layers, each layer of multi-scale feature fusion layer fuses the coding features of the corresponding downsampling layer in the encoder, the fusion features of the previous layer of multi-scale feature fusion layer and the attention features output by the attention module. The present invention fully considers global features and local features, effectively improves the detection accuracy of the drivable area, and at the same time improves the inference speed of the model, thereby achieving the effect of real-time detection.

[0065] Example 2

[0066] An embodiment of the present invention provides a domain controller, including:

[0067] One or more processors for deploying a drivable area detection model;

[0068] a memory for storing one or more computer programs,

[0069] When the one or more computer programs are executed by the one or more processors, the one or more processors implement the drivable area detection method as described in any one of the first embodiments.

[0070] Figure 4 In the example, a processor 10 is used; the processor 10 and the memory 11 in the domain controller can be connected by a bus or other means. Figure 4 The bus connection is taken as an example.

[0071] Exemplarily, the computer program can be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor 10 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments that can perform specific functions, and the instruction segments are used to describe the execution process of the computer program in the domain controller. For example, the computer program can be divided into an image acquisition module for acquiring a road surface image; a road surface segmentation module for inputting the road surface image into a drivable area detection model, and performing road surface segmentation on the road surface image to identify the drivable area in the road surface image, that is, the drivable area detection method described in Example 1.

[0072] The processor 10 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The domain controller is the control center of the vehicle and controls the vehicle to perform autonomous driving based on the identified drivable area.

[0073] The memory 11 can be used to store computer programs and / or modules, such as the program instructions / modules corresponding to the obstacle target detection method in Example 1 of the present invention. The processor 10 implements the various functions of the domain controller by running or executing the computer programs and / or modules stored in the memory 11 and accessing the data stored in the memory 11. The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system, at least one application required for a function, and the data storage area may store data generated based on the use of a terminal (such as a vehicle). In addition, the memory 11 may include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device. In some examples, the memory 11 may further include a memory remotely located from the processor 10, and such remote memory may be connected to the domain controller via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0074] It should be noted that the principles and technical effects of the embodiment of the present invention are the same as those of the drivable area detection method described in Example 1, and will not be described in detail here.

[0075] Example 3

[0076] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program. When the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the drivable area detection method as described in any one of the embodiments.

[0077] The drivable area detection method includes:

[0078] Acquire road surface images;

[0079] Inputting the road surface image into a drivable area detection model, performing road surface segmentation on the road surface image to identify the drivable area in the road surface image;

[0080] The drivable area detection model includes an encoder, a decoder, and an attention module. The attention module extracts attention features from the encoding features output by the encoder and inputs them into the decoder for feature fusion.

[0081] Among them, the encoder adopts a residual network structure, including N sequentially connected downsampling layers; the decoder includes N sequentially connected multi-scale feature fusion layers, and is connected one-to-one with the downsampling layers. Each layer of multi-scale feature fusion layer fuses the encoding features of the corresponding downsampling layer in the encoder, the fusion features of the previous multi-scale feature fusion layer, and the attention features output by the attention module.

[0082] Of course, the computer-readable storage medium provided in the embodiment of the present invention is not limited to the method operations described above, and can also perform related operations in the drivable area detection method provided in any embodiment of the present invention.

[0083] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general-purpose hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0084] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting a drivable area, characterized in that: include: Acquire road surface images; Inputting the road surface image into a drivable area detection model, performing road segmentation on the road surface image to identify the drivable area in the road surface image; The drivable area detection model includes an encoder, a decoder, and an attention module. The attention module extracts attention features from the encoding features output by the encoder and inputs them into the decoder for feature fusion. The encoder adopts a residual network structure, including N sequentially connected downsampling layers; the decoder includes N sequentially connected multi-scale feature fusion layers, and is connected one-to-one with the downsampling layers. Each multi-scale feature fusion layer fuses the encoding features of the corresponding downsampling layer in the encoder, the fusion features of the previous multi-scale feature fusion layer, and the attention features output by the attention module; an edge enhancement module is provided between some of the downsampling layers and the multi-scale feature fusion layers to extract edge features; the edge enhancement module uses an HPF operator to extract edge features; The attention module includes a first stage, a second stage, a convolutional neural network layer, a multi-head self-attention mechanism layer and a multi-layer perceptron; wherein the input of the first stage is the encoded feature output by the Nth downsampling layer; the input of the convolutional neural network layer and the multi-head self-attention mechanism layer is the feature output by the first stage; the input of the second stage is the concatenation operation result of the encoded feature output by the Nth downsampling layer and the feature output by the convolutional neural network layer and the multi-head self-attention mechanism layer; the input of the multi-layer perceptron is the feature output by the second stage; the output of the attention module is the concatenation operation result of the feature input to the second stage and the feature output by the multi-layer perceptron; the first stage includes a first feature slicing layer, a first batch normalization layer, a first instance normalization layer and a first feature splicing layer; wherein the input of the first feature slicing layer is the encoded feature output by the Nth downsampling layer, the input of the first batch normalization layer and the first instance normalization layer is the feature output by the first feature slicing layer, the input of the first feature splicing layer is the feature output by the first batch normalization layer and the first instance normalization layer, and the output of the first feature splicing layer is the output of the first stage.

2. The method for detecting a drivable area according to claim 1, wherein: The residual network structure adopted by the encoder is ResNet34; wherein, the number of layers of the downsampling layer and the multi-scale feature fusion layer are both 4.

3. The method for detecting a drivable area according to claim 1, wherein: The input of the multi-scale feature fusion layer includes: the concatenation of the encoding features output by the corresponding downsampling layer and the edge features output by its corresponding edge enhancement module, the fusion features output by the previous multi-scale feature fusion layer, and the attention features output by the attention module.

4. The method for detecting a drivable area according to claim 1, wherein: The template of the HPF operator is:

5. The method for detecting a drivable area according to claim 1, wherein: The function of the multi-scale feature fusion layer is expressed as follows: Among them, α4+α5=1, α1+α2+α3=1, f up Represents upsampling operation, MSFF i Represents the features output by the i-th multi-scale feature fusion layer, MSFF i-1 represents the features output by the i-1th multi-scale feature fusion layer, f conv1×1 represents a convolution layer with a convolution kernel size of 1, GLIBN is the attention feature output by the attention module, It is the encoded feature output by the 4-i-1th downsampling layer.

6. The method for detecting a drivable area according to claim 5, wherein: The second stage includes a second feature slice layer, a second batch normalization layer, a second instance normalization layer and a second feature splicing layer; wherein, the input of the second feature slice layer is the input of the second stage, the input of the second batch normalization layer and the second instance normalization layer is the features output by the second feature slice layer, the input of the second feature splicing layer is the features output by the second batch normalization layer and the second instance normalization layer, and the output of the second feature splicing layer is the output of the second stage.

7. A domain controller, characterized in that: The domain controller includes: One or more processors for deploying a drivable area detection model; a memory for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the drivable area detection method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the drivable area detection method according to any one of claims 1 to 6.

Images