Feature extraction method and convolutional neural network based on automatic parking model

Abstract

The present invention relates to a feature extraction method, convolutional neural network, on-board computing system, and computer storage medium based on an automated parking model. The method comprises: A. feeding a captured image of the parking environment into the automated parking model as input information; B. extracting features from the input information using the automated parking model's backbone network. The backbone network comprises four feature extraction stages, each of which includes a series of convolutional layers with kernel sizes of 1*3, 3*3, and 3*1, resulting in a larger receptive field than an equivalent convolutional layer structure with a kernel size of 5*5.

Description

Technical Field

[0001] The present invention relates to the field of automatic parking technology, and more particularly to a feature extraction method based on an automatic parking model, a convolutional neural network, an on-vehicle computing system, and a computer storage medium. Background Art

[0002] The application and deployment of convolutional neural networks (CNNs) has always been a challenge for autonomous driving perception algorithms, especially those for automated parking. Current mainstream CNNs typically utilize complex operations, including attention mechanisms, grouped convolutions, and custom calculations, to improve model performance. These complex operations significantly limit their deployment on-board. Limited by the inability of onboard computing platforms to support specialized operations within the model, coupled with extremely high speed and precision requirements, the design of an efficient and easily deployable backbone network has become a pressing challenge in the automated parking field. Summary of the Invention

[0003] In order to solve or at least alleviate one or more of the above problems, the following technical solutions are provided.

[0004] According to a first aspect of the present invention, a feature extraction method based on an automatic parking model is provided, comprising: A. inputting a captured parking environment image as input information into the automatic parking model; B. performing feature extraction on the input information using a backbone network of the automatic parking model; wherein the backbone network includes four feature extraction stages, each of the four feature extraction stages including a convolutional layer structure having convolution kernel sizes of 1*3, 3*3, and 3*1, connected in series, to achieve a receptive field equivalent to a convolutional layer structure having a convolution kernel size of 5*5.

[0005] As an alternative or supplement to the above scheme, in a method according to an embodiment of the present invention, the four feature extraction stages include a first extraction stage, a second extraction stage, a third extraction stage and a fourth extraction stage, and each of the four feature extraction stages includes: a first convolution block with an equivalent step size of 2; and one or more second convolution blocks with an equivalent step size of 1.

[0006] As an alternative or supplement to the above scheme, in a method according to an embodiment of the present invention, the following processing operations are performed on the input information in the first convolution block: the input data is copied into first input data and second input data; the first input data is input into the first convolution layer to generate first output data, wherein the step size of the first convolution layer is 2 and the convolution kernel size is 3*3; the second input data is input into the second convolution layer, the third convolution layer, and the fourth convolution layer connected in series to generate second output data, wherein the step size of the second convolution layer is 1 and the convolution kernel size is 1*3, the step size of the third convolution layer is 2 and the convolution kernel size is 3*3, and the step size of the fourth convolution layer is 1 and the convolution kernel size is 3*1; and the first output data and the second output data are spliced ​​in the channel dimension to generate third output data.

[0007] As an alternative or supplement to the above scheme, in a method according to an embodiment of the present invention, the following processing operations are performed on the input information in the second convolution block: the third output data is input into the fifth convolution layer to generate fourth output data, wherein the stride of the fifth convolution layer is 1 and the convolution kernel size is 1*1; the fourth output data is divided into two parts according to the channel dimension to generate fifth output data and sixth output data, wherein the number of channels of the fifth output data and the sixth output data is half of the number of channels of the fourth output data; the sixth output data is input into the sixth convolution layer, the seventh convolution layer, and the eighth convolution layer connected in series to generate seventh output data, wherein the stride of the sixth convolution layer is 1 and the convolution kernel size is 1*3, the stride of the seventh convolution layer is 1 and the convolution kernel size is 3*3, and the stride of the eighth convolution layer is 1 and the convolution kernel size is 3*1; and the fifth output data and the seventh output data are spliced ​​in the channel dimension to generate eighth output data.

[0008] As an alternative or supplement to the above scheme, in a method according to one embodiment of the present invention, the number of convolution kernels of the sixth convolution layer is twice the number of channels of the sixth output data, the number of convolution kernels of the seventh convolution layer is equal to the number of convolution kernels of the sixth convolution layer, and the number of convolution kernels of the eighth convolution layer is half the number of convolution kernels of the sixth convolution layer.

[0009] As an alternative or supplement to the above scheme, in a method according to one embodiment of the present invention, the first extraction stage and the second extraction stage only include one second convolution block, and the number of second convolution blocks included in the third extraction stage and the fourth extraction stage depends on the model computing power and the expected inference speed.

[0010] As an alternative or supplement to the above scheme, in a method according to one embodiment of the present invention, the number of input channels of the first extraction stage is N, the number of input channels of the second extraction stage is 2*N, the number of input channels of the third extraction stage is 4*N, and the number of input channels of the fourth extraction stage is 8*N, where the size of N depends on the model computing power and the expected inference speed.

[0011] As an alternative or supplement to the above scheme, in a method according to an embodiment of the present invention, the backbone network also includes a fast downsampling stage, in which the following processing operations are performed on the input information: the input data is input into a convolution layer with a step size of 2, a void convolution rate of 1, and a convolution kernel size of 3*3.

[0012] According to a second aspect of the present invention, a convolutional neural network is provided, comprising a backbone network for performing feature extraction on a parking environment image. The backbone network comprises four feature extraction stages, each of the four feature extraction stages comprising a convolutional layer structure having convolution kernel sizes of 1*3, 3*3, and 3*1, connected in series, with a receptive field equivalent to a convolutional layer structure having a convolution kernel size of 5*5.

[0013] According to a third aspect of the present invention, there is provided an in-vehicle computing system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any one of the methods described in the first aspect of the present invention is implemented.

[0014] According to a fourth aspect of the present invention, there is provided a computer storage medium, the computer storage medium comprising instructions, the instructions being used to execute any one of the methods according to the first aspect of the present invention when run.

[0015] On the one hand, the feature extraction backbone network according to one or more embodiments of the present invention includes four feature extraction stages, each of which is composed of a simple basic convolution operation block. Therefore, it can be directly deployed on the on-board computing platform without additional processing, solving the problem of limited deployment on the vehicle side.

[0016] On the other hand, according to one or more embodiments of the present invention, by using low-cost equivalent replacement of large receptive fields (for example, using convolutional layer structures with convolution kernel sizes of 1*3, 3*3, and 3*1 to approximate the equivalent convolution layer with a convolution kernel size of 5*5), channel multiplexing operations (for example, dividing the input data of the second convolution block into two parts according to the channel dimension) and channel expansion technology, the model performance is effectively improved while achieving the fast reasoning capability of the convolutional neural network, thereby meeting the requirements of fast speed, high performance and easy deployment for the vehicle-side parking model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and / or other aspects and advantages of the present invention will become clearer and easier to understand through the following description of various aspects in conjunction with the accompanying drawings, in which the same or similar elements are represented by the same reference numerals. In the accompanying drawings:

[0018] Figure 1 is a flowchart of a feature extraction method 10 based on an automatic parking model according to one or more embodiments of the present invention;

[0019] Figure 2 is a schematic block diagram of a backbone network 20 according to one or more embodiments of the present invention;

[0020] Figure 3 is a schematic block diagram of a first convolution block according to one or more embodiments of the present invention;

[0021] Figure 4 is a schematic block diagram of a second convolution block according to one or more embodiments of the present invention; and

[0022] Figure 5 FIG. 5 is a block diagram of an onboard computing system 50 according to one embodiment of the present invention. DETAILED DESCRIPTION

[0023] The description of the following specific embodiments is merely exemplary in nature and is not intended to limit the disclosed technology or the application and use of the disclosed technology. In addition, there is no intention to be bound by any express or implied theory presented in the foregoing technical field, background technology or the following specific embodiments.

[0024] In the following detailed description of the embodiments, numerous specific details are set forth to provide a more thorough understanding of the disclosed technology. However, it will be apparent to one of ordinary skill in the art that the disclosed technology can be practiced without these specific details. In other instances, well-known features are not described in detail to avoid unnecessarily complicating the description.

[0025] Terms such as "comprising" and "including" indicate that, in addition to the units and steps directly and explicitly stated in the specification, the technical solution of the present invention does not exclude the presence of other units and steps not directly or explicitly stated. Terms such as "first" and "second" do not indicate the order of units in terms of time, space, size, etc., but are merely used to distinguish between units.

[0026] It should be noted that, in the context of the present invention, the term "receptive field" refers to the size of the original image corresponding to a certain neural unit within a neural convolutional network, wherein the larger the receptive field, the larger the range of the original image corresponding to the neural unit, which means that it may contain more global and higher semantic features. The term "convolution kernel size" refers to the size range of the convolution, which represents the size of the receptive field in a convolutional neural network. Generally speaking, the larger the convolution kernel, the larger the receptive field. The term "step size" refers to the length spanned by each convolution when the convolution kernel performs a convolution operation on the image.

[0027] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

[0028] Figure 1 FIG. 1 is a flowchart of a feature extraction method 10 based on an automatic parking model according to one or more embodiments of the present invention.

[0029] like Figure 1 As shown, in step S110, the collected parking environment image is input into the automatic parking model as input information.

[0030] Exemplarily, the parking environment image is an RGB image, i.e., an image consisting of a three-dimensional array in the format of M*N*3, where R represents red, G represents green, and B represents blue. Exemplarily, the automated parking model is a convolutional neural network (CNN) model.

[0031] In step S120 , the backbone network of the automatic parking model is used to extract features from the input information.

[0032] The backbone network refers to the network portion of a convolutional neural network model that can be used to extract features. It can connect different modules in the network. The backbone network can automatically extract useful feature information through training. The backbone networks in some classic networks such as Resnet and VGG are often used as references by other neural networks because these backbone networks have strong feature extraction capabilities in classification problems. It is important to note that the present invention creatively applies a new backbone network to the automatic parking model, and through this new backbone network, it meets the requirements of the vehicle-side parking model for fast speed, high performance, and easy deployment.

[0033] The following will continue to refer to Figure 2-4 Let's introduce the specific architecture of the backbone network in detail.

[0034] Figure 2 FIG. 2 is a schematic block diagram of a backbone network 20 according to one or more embodiments of the present invention.

[0035] Optionally, backbone network 20 includes a fast downsampling stage. This fast downsampling module is used to reduce the resolution of the input image and simultaneously convert the input image from the image domain to the feature domain. The following describes the various stages of the feature extraction process, including the convolutional blocks and data processing operations in the convolutional layers, using an input RGB image of size 384*480*3 as an example.

[0036] Optionally, the fast downsampling stage may include a convolution layer with a stride of 2, a dilated convolution rate of 1, a convolution kernel size of 3*3, and a number of convolution kernels of 32. For example, after the fast downsampling stage, an input image of size 384*480*3 obtains a feature map of size approximately 192*240*32, which is input as input data to a subsequent processing stage.

[0037] like Figure 2 As shown, backbone network 20 includes four feature extraction stages, each of which is used to extract discriminative features at different levels from the input data. For example, the four feature extraction stages are sequentially labeled as stage 1, stage 2, stage 3, and stage 4, in the order of data flow.

[0038] Optionally, each feature extraction stage includes a first convolution block (block 1) with an equivalent stride of 2 and one or more second convolution blocks (block 2) with an equivalent stride of 1. Furthermore, the first extraction stage and the second extraction stage include only one first convolution block and one second convolution block, while the third extraction stage and the fourth extraction stage may include one first convolution block and multiple second convolution blocks, where the number of second convolution blocks included in the third extraction stage and the fourth extraction stage depends on the model computing power and the expected inference speed.

[0039] It should be noted that since the resolution of the data processing objects in the first extraction stage and the second extraction stage is usually greater than that in the third extraction stage and the fourth extraction stage, designing the first extraction stage and the second extraction stage as a simpler structure (i.e., including only 1 block 1 and 1 block 2) will help reduce the amount of inference calculation while maintaining network performance.

[0040] Optionally, the number of input channels in the first extraction stage is set to N, the number of input channels in the second extraction stage is set to 2*N, the number of input channels in the third extraction stage is set to 4*N, and the number of input channels in the fourth extraction stage is set to 8*N, where N is a positive integer and the size of N depends on the model computing power and the expected inference speed. For example, N is equal to 32.

[0041] Optionally, each feature extraction stage includes a convolution layer structure with convolution kernel sizes of 1*3, 3*3, and 3*1 connected in series in sequence to approximate a convolution structure with an equivalent convolution kernel size of 5*5. Exemplarily, the first convolution block and the second convolution block in each feature extraction stage include the above-mentioned convolution layer structures of 1*3, 3*3, and 3*1. The advantages of using multiple smaller convolution kernels to be equivalent to a larger convolution kernel are: first, when obtaining the same receptive field (for example, the receptive field of 1*3, 3*3, 3*1 is the same as the receptive field of 5*5), the computational complexity of multiple smaller convolution kernels is small, thereby improving the computational speed of the model; second, compared with a larger convolution kernel, multiple smaller convolution kernels can introduce more nonlinearity, thereby improving the depth of the neural convolution network and improving the performance of the neural network to a certain extent.

[0042] The following will be combined Figure 3 and Figure 4 The specific structure and processing process of the first convolution block and the second convolution block are introduced in detail.

[0043] Figure 3 FIG. 1 is a schematic block diagram of a first convolution block according to one or more embodiments of the present invention. Figure 3 The first convolutional block (block 1) shown has an equivalent stride of 2.

[0044] As described above, a feature map of approximately 192*240*32 is input as input data to the first convolution block (block 1) in the first extraction stage (stage 1). That is, the number of input channels in the first extraction stage is 32. It should be noted that the data size described herein can be set according to actual needs and is not limited to the numerical range shown in this embodiment.

[0045] In the first convolution block, the input data (tensor 0) of size 192*240*32 is first copied to generate the first input data (tensor 1) and the second input data (tensor 2) of size 192*240*32. Then, different convolution operations are performed on the first input data and the second input data.

[0046] Optionally, the first input data is input to the first convolutional layer to generate first output data (tensor 3), where the stride of the first convolutional layer is 2 and the convolution kernel size is 3*3. For example, the first input data of size 192*240*32 passes through the first convolutional layer (for example, the number of convolution kernels is 32), and the output data size is approximately 91*120*32.

[0047] Optionally, the second input data is fed into a second, third, and fourth convolutional layer, connected in series, to generate second output data (tensor 4). The stride of the second convolutional layer is 1, and the convolution kernel size is 1*3; the stride of the third convolutional layer is 2, and the convolution kernel size is 3*3; and the stride of the fourth convolutional layer is 1, and the convolution kernel size is 3*1. For example, data of size 192*240*32 passes through the second convolutional layer (e.g., with 32 convolution kernels), and the output data size is approximately 192*240*32. For example, data of size 192*240*32 passes through the third convolutional layer (e.g., with 32 convolution kernels), and the output data size is approximately 91*120*32. For example, data of size 91*120*32 passes through the fourth convolutional layer (e.g., with 32 convolution kernels), and the output data size is approximately 91*120*32.

[0048] Optionally, the first output data (tensor 3) and the second output data (tensor 4) are concatenated along the channel dimension to generate third output data (tensor 5). For example, the concatenated data size of the first and second output data, which is 91*120*32, is approximately 91*120*64. Next, the third output data (tensor 5) is input into the second convolutional block (block 2) in the first extraction stage (stage 1).

[0049] Figure 4 FIG. 1 is a schematic block diagram of a second convolution block according to one or more embodiments of the present invention. Figure 4 The second convolutional block (block 2) shown has an equivalent stride of 1.

[0050] As mentioned above, the feature map of approximately 91*120*64 is input as input data to the second convolution block (block 2) in the first extraction stage (stage 1).

[0051] In the second convolutional block, the third output data (tensor 5) is first input to the fifth convolutional layer to generate the fourth output data (tensor 6). The stride of the fifth convolutional layer is 1, and the convolution kernel size is 1*1. For example, data of size 91*120*64 passes through the fifth convolutional layer (for example, with 64 convolution kernels), and the output data size is approximately 91*120*64. The significance of introducing a 1*1 convolution kernel here is to enable sufficient information exchange between the first output data and the second output data contained in the third output data, which have undergone different convolution operations, thereby enhancing the network's expressive power.

[0052] Optionally, the fourth output data (tensor 6) is divided into two parts according to the channel dimension to generate fifth output data (tensor 7) and sixth output data (tensor 8), where the number of channels of the fifth and sixth output data is half of the number of channels of the fourth output data. For example, data of size 91*120*64 is divided into two feature maps of size 91*120*32.

[0053] Optionally, the sixth output data (tensor 8) is input into the sixth convolutional layer, the seventh convolutional layer, and the eighth convolutional layer connected in series to generate the seventh output data (tensor 9), where the stride of the sixth convolutional layer is 1 and the convolution kernel size is 1*3, the stride of the seventh convolutional layer is 1 and the convolution kernel size is 3*3, and the stride of the eighth convolutional layer is 1 and the convolution kernel size is 3*1.

[0054] Optionally, channel expansion techniques are introduced in the sixth, seventh, and eighth convolutional layers to further improve model performance. Exemplarily, the number of convolution kernels in the sixth convolutional layer is twice the number of input channels of the sixth output data, the number of convolution kernels in the seventh convolutional layer is equal to the number of convolution kernels in the sixth convolutional layer, and the number of convolution kernels in the eighth convolutional layer is half the number of convolution kernels in the sixth convolutional layer. Exemplarily, data of size 91*120*32 passes through the sixth convolutional layer (e.g., with 64 convolution kernels), resulting in an output data size of approximately 91*120*64. Exemplarily, data of size 91*120*64 passes through the seventh convolutional layer (e.g., with 64 convolution kernels), resulting in an output data size of approximately 91*120*64. Exemplarily, data of size 91*120*64 passes through the eighth convolutional layer (e.g., with 32 convolution kernels), resulting in an output data size of approximately 91*120*32.

[0055] Optionally, the fifth output data (tensor 7) and the seventh output data (tensor 9) are concatenated along the channel dimension to generate the eighth output data (tensor 10). For example, the concatenated data size of the fifth and seventh output data, which are 91*120*32, is approximately 91*120*64. Next, the eighth output data (tensor 10) is input into the first convolutional block (block 1) in the second extraction stage (stage 2).

[0056] Since the convolutional layer structures in the first convolutional block and the second convolutional block in the second, third, and fourth extraction stages are the same as those in the first extraction stage (except that the number of input channels changes according to the rules described above), they will not be described in detail.

[0057] Figure 5FIG. 5 is a block diagram of a vehicle-mounted computing system 50 according to an embodiment of the present invention. Figure 5 As shown in FIG, the computer device 50 includes a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 530, the following is achieved: Figure 1 The method shown.

[0058] In addition, as described above, the present invention can also be implemented as a computer storage medium, in which a program for causing a computer to execute the following steps is stored. Figure 1 The procedure of the method shown.

[0059] Here, as computer storage media, various types of computer storage media can be used, such as disks (for example, magnetic disks, optical disks, etc.), cards (for example, memory cards, optical cards, etc.), semiconductor memories (for example, ROMs, non-volatile memories, etc.), and tapes (for example, magnetic tapes, cassette tapes, etc.).

[0060] First, according to some embodiments of the present invention, the feature extraction backbone network is composed of simple basic convolution operation blocks, so it can be directly deployed on the vehicle computing platform without additional processing, solving the problem of limited deployment of convolutional neural networks on the vehicle side.

[0061] Secondly, some embodiments of the present invention effectively improve the model performance while achieving the fast reasoning capability of convolutional neural networks by using low-cost equivalent replacements for large receptive fields, channel multiplexing, and channel expansion technologies, thereby meeting the requirements for fast speed, high performance, and easy deployment of vehicle-side parking models.

[0062] In the case of applicable, hardware, software or a combination of hardware and software can be used to realize the various embodiments provided by the present invention. Moreover, in the case of applicable, without departing from the scope of the present invention, the various hardware components and / or software components set forth herein can be combined into composite components comprising software, hardware and / or both. In the case of applicable, without departing from the scope of the present invention, the various hardware components and / or software components set forth herein can be divided into subcomponents comprising software, hardware or both. In addition, in the case of applicable, it is contemplated that software components can be implemented as hardware components, and vice versa.

[0063] Software according to the present invention (such as program code and / or data) can be stored on one or more computer storage media. It is also contemplated that the software identified herein can be implemented using one or more general-purpose or special-purpose computers and / or computer systems, networked and / or otherwise. Where applicable, the order of the various steps described herein can be changed, combined into composite steps, and / or divided into sub-steps to provide the features described herein.

[0064] The embodiments and examples set forth herein are provided to best illustrate embodiments according to the present invention and its specific applications, and thereby enable those skilled in the art to make and use the invention. However, those skilled in the art will appreciate that the above description and examples are provided for ease of illustration and example only. The descriptions set forth are not intended to encompass all aspects of the invention or to limit the invention to the precise forms disclosed.

Claims

1. A feature extraction method based on an automatic parking model, characterized in that: include: A. inputting the collected parking environment image as input information into the automatic parking model; B. extracting features from the input information using the backbone network of the automatic parking model; The backbone network includes four feature extraction stages, each of which includes a convolutional layer structure with convolution kernel sizes of 1*3, 3*3, and 3*1 connected in series, with a receptive field of a convolutional layer structure with an equivalent convolution kernel size of 5*5. Each of the four feature extraction stages includes a first convolution block with an equivalent step size of 2, and the following processing operations are performed on the input information in the first convolution block: Copying the input data into first input data and second input data; Inputting the first input data into a first convolutional layer to generate first output data, wherein the stride of the first convolutional layer is 2 and the convolution kernel size is 3*3; Inputting the second input data into a second convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in series to generate second output data, wherein the stride of the second convolutional layer is 1 and the convolution kernel size is 1*3, the stride of the third convolutional layer is 2 and the convolution kernel size is 3*3, and the stride of the fourth convolutional layer is 1 and the convolution kernel size is 3*1; and The first output data and the second output data are concatenated in a channel dimension to generate third output data.

2. The method according to claim 1, wherein The four feature extraction stages include a first extraction stage, a second extraction stage, a third extraction stage and a fourth extraction stage, and each of the four feature extraction stages includes one or more second convolution blocks with an equivalent step size of 1.

3. The method according to claim 2, wherein: In the second convolution block, the following processing operations are performed on the input information: Inputting the third output data into a fifth convolutional layer to generate fourth output data, wherein the stride of the fifth convolutional layer is 1 and the convolution kernel size is 1*1; Dividing the fourth output data into two parts according to a channel dimension to generate fifth output data and sixth output data, wherein the number of channels of the fifth output data and the sixth output data is half of the number of channels of the fourth output data; The sixth output data is input into the sixth convolutional layer, the seventh convolutional layer, and the eighth convolutional layer connected in series to generate the seventh output data, wherein the step size of the sixth convolutional layer is 1, and the convolution kernel size is 1*3, the step size of the seventh convolutional layer is 1, and the convolution kernel size is 3*3, and the step size of the eighth convolutional layer is 1, and the convolution kernel size is 3*1; and the fifth output data and the seventh output data are spliced ​​in the channel dimension to generate the eighth output data.

4. The method according to claim 3, wherein: The number of convolution kernels of the sixth convolution layer is twice the number of channels of the sixth output data, the number of convolution kernels of the seventh convolution layer is equal to the number of convolution kernels of the sixth convolution layer, and the number of convolution kernels of the eighth convolution layer is half the number of convolution kernels of the sixth convolution layer.

5. The method according to claim 2, wherein: The first extraction stage and the second extraction stage include only one second convolution block, and the number of second convolution blocks included in the third extraction stage and the fourth extraction stage depends on the model computing power and the expected inference speed.

6. The method according to claim 2, wherein: The number of input channels in the first extraction stage is N, the number of input channels in the second extraction stage is 2*N, the number of input channels in the third extraction stage is 4*N, and the number of input channels in the fourth extraction stage is 8*N, where the size of N depends on the model computing power and the expected inference speed.

7. The method according to claim 1, wherein The backbone network also includes a fast downsampling stage, in which the following processing operations are performed on the input information: the input data is input into a convolution layer with a step size of 2, a void convolution rate of 1, and a convolution kernel size of 3*3.

8. A method for constructing a convolutional neural network, characterized in that The convolutional neural network includes a backbone network, which is used to extract features from parking environment images. The backbone network includes four feature extraction stages, each of which includes a convolution layer structure with convolution kernel sizes of 1*3, 3*3, and 3*1 connected in series, with a receptive field of a convolution layer structure with an equivalent convolution kernel size of 5*5. Each of the four feature extraction stages includes a first convolution block with an equivalent step size of 2, and in the first convolution block: The input data is copied into first input data and second input data; The first input data is input into a first convolutional layer to generate first output data, wherein the stride of the first convolutional layer is 2 and the convolution kernel size is 3*3; The second input data is input into a second convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in series to generate second output data, wherein the stride of the second convolutional layer is 1 and the convolution kernel size is 1*3, the stride of the third convolutional layer is 2 and the convolution kernel size is 3*3, and the stride of the fourth convolutional layer is 1 and the convolution kernel size is 3*1; and The first output data and the second output data are concatenated in a channel dimension to generate third output data.

9. An in-vehicle computing system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

10. A computer storage medium, characterized in that The computer storage medium comprises instructions that, when executed, perform the method according to any one of claims 1 to 7.

Images