Real-time binocular stereo matching method and automobile suspension control system based on the method

Abstract

The application provides a real-time binocular stereo matching method and a vehicle suspension control system based on the method. The method comprises the following steps: obtaining multi-scale features of left and right views based on a feature extraction network, the features being represented by a plurality of three-dimensional tensors, thereby constructing three fusion cost volumes (texture volume, correlation volume and mutual volume) based on low-level feature maps with 1 / 4 resolution to realize cost volume regression, then realizing cost regression by a sandglass-shaped stacked 3D convolution combined with other level feature maps, finally performing disparity regression at 1 / 4 resolution and original resolution and obtaining a disparity map at the original resolution through supervised learning. Based on the obtained disparity map and a parameter matrix of a binocular camera, a depth map of a road surface can be calculated, and then by comparing the depth difference between the front and rear frames, the change of the road surface relief can be calculated by using the geometric relationship, and then the change is used as an input to perform hierarchical control of the suspension, so as to ensure driving comfort.

Description

Technical Field

[0001] This invention relates to the field of automotive technology, specifically to a real-time binocular stereo matching method and a vehicle suspension control system based on the method. Background Technology

[0002] In the field of intelligent driving, binocular stereo matching is a crucial component of environmental perception technology and holds significant importance for vehicle suspension control. Specifically, it uses a pair of calibrated binocular cameras to capture environmental images, simulating human visual perception, thereby accurately locating the corresponding pixels in the right view and the left view. In suspension control tasks, accurately predicting road surface undulations is of significant guiding importance for active suspension control.

[0003] In recent years, many researchers have devoted themselves to the study of binocular stereo matching, aiming to find efficient and accurate matching methods. Current stereo matching methods mainly consist of four modules: shared-weight feature extraction, cost volume construction, cost volume aggregation, and disparity regression. Among them, cost volume construction encodes the feature map into a 4D tensor containing left and right view similarity information, providing a coarse representation for matching pixel pairs. Therefore, the richness and accuracy of the cost volume information constructed in this process basically determine the final stereo matching performance.

[0004] However, existing construction strategies only consider the correlation between the left and right views, neglecting the problem of blurred similarity information caused by ill-posed regions formed by strong light reflection and large-area repetition in some areas. This greatly limits the model's ability to accurately model complex regions. In addition, existing strategies only consider the matching cost when there is a one-to-one correspondence between pixels in the left and right feature maps, ignoring two key issues: first, when the left and right feature maps are used as disparity probability distributions, the similarity information between their overall distributions is extremely important for cost volume construction; second, due to different shooting angles, a single pixel in the left view may correspond to multiple pixels in the right view.

[0005] Meanwhile, the automotive suspension system is crucial to vehicle performance and ride comfort. It not only affects the vehicle's stability and handling under different road conditions but also the passenger's comfort experience. High-quality suspension control methods can effectively absorb road impacts, reduce the amplitude of vibrations transmitted to the seat, and improve driving safety and comfort. Furthermore, with the development of automotive technology, the intelligent and automated nature of suspension systems is becoming an important future trend, and adjusting the suspension posture according to real-time road conditions to achieve optimal driving performance is a current key research direction.

[0006] Despite advancements in existing intelligent suspension control technology, shortcomings remain. While in-vehicle cameras are widely used in various vehicle models, current control strategies largely fail to fully utilize advanced environmental perception technologies. They rely primarily on vehicle sensors to indirectly perform active or semi-active control based on changes in the vehicle's attitude.

[0007] A search of patent literature revealed that invention patent CN202410355410.1 discloses a binocular depth estimation method based on automatic network compression. This method acquires the left and right views and their corresponding disparity maps, constructing a binocular depth estimation model including NASCNet and NASSNet. The model training employs a progressive search strategy, first training the network with the largest structure, and then automatically searching for the convolutional kernel size, model depth, and width to better fit the data. However, this method has limitations in optimizing the cost-volume construction strategy, resulting in limited matching accuracy.

[0008] In summary, given the problems of the existing technologies, researching a real-time binocular stereo matching method and a method-based automotive suspension control system has become a critical task that urgently needs to be addressed. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the purpose of this invention is to provide a real-time binocular stereo matching method and a vehicle suspension control system based on the method.

[0010] A real-time binocular stereo matching method based on three-way fusion cost volume provided by the present invention includes the following steps:

[0011] S1: By processing the original binocular image, multi-level feature maps in the form of three-dimensional tensors at different scales are obtained to acquire semantic and texture information at different levels;

[0012] S2: Based on the original binocular image and the low-level feature map of the multi-level feature map, a three-fusion cost volume is constructed to construct the similarity and correlation of the left and right views in the depth direction.

[0013] S3: Based on the three-fusion cost volume and other layer-level feature maps of the multi-level feature maps, the cost volume is aggregated by performing downsampling-upsampling layer by layer according to the tensor size, based on 3D convolution, to obtain the aggregated cost volume;

[0014] S4: Input the aggregated cost volume into the dual-scale disparity regression module, and use the arg-softmax function to output disparity maps at 1 / 4 resolution and original resolution, respectively.

[0015] S5: Using the parameters of the binocular camera and the disparity map, the depth map of the road surface is calculated. By comparing the depth changes in the depth map in consecutive frames, the change in road surface undulation is calculated based on the geometric relationship, which serves as the input signal for subsequent suspension control.

[0016] Preferably, step S1 includes the following sub-steps:

[0017] Step S1.1: Regularize the original binocular image and expand it into a four-dimensional tensor to obtain the processed image;

[0018] Step S1.2: The processed image is passed through a lightweight feature extraction network, and initial multi-scale feature maps are obtained by progressively using convolution, pooling, and activation layers. ;

[0019] Step S1.3, based on the initial multi-scale feature map, and using the structure of the feature pyramid, perform the following operations sequentially: ① For After upsampling and By splicing and combining, we obtain ; ② Further encoding using 1×1 convolutions yields new values ​​at the current level. ③ Thus, multi-level feature maps are obtained. .

[0020] Preferably, in step S1.1, the regularization factors applied to the original binocular image are as follows: Where R, G, and B represent the three channels of the image, respectively. This represents the mean. Indicates variance.

[0021] Preferably, in step S1.3, the multi-level feature map These correspond to feature information at resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32, respectively, and the size of their feature dimensions corresponds to the maximum disparity value that a binocular camera can capture. Relatedly, the feature map with 1 / 4 resolution is a low-level feature map, and the feature maps with the remaining resolution are feature maps of other levels.

[0022] Preferably, step S2 includes the following sub-steps:

[0023] Step S2.1, based on the original stereo images The algorithm uses a shared-weight module to extract texture information at the original resolution, and then designs a cascaded, cross-layer connected convolutional neural network to encode the texture information at the original resolution, thereby extracting the texture information at the low resolution. ,Will With low-level feature maps Aggregation, using 3D convolution to calculate texture volume The calculation formula is as follows:

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] in, and These are all intermediate outputs of the network. This indicates a normalization operation. Represents two-dimensional convolution. Represents 3D convolution. Represents the Hadamard product of matrices. This represents the tensor concatenation operation. This represents a compound operation of addition, convolution, and regularization.

[0030] Step S2.2, based on low-level feature maps First, use cosine similarity to roughly calculate the correlation tensor between the two. Then, using a deformable attention mechanism, for The location of each pixel in eigenvectors It adaptively searches for and matches other relevant pixels, and after weighted aggregation of the feature information of each relevant pixel, obtains a new feature vector. Repeat this operation for each position to obtain a new correlation tensor. The correlation volume is calculated using 3D convolution. The calculation formula is as follows:

[0031]

[0032] in, Linear Represents a linear layer. For correlation tensor The intermediate layer output obtained after further learning through two-dimensional convolution. This represents the bias in adaptive learning. This represents the aggregated weights after adaptive learning;

[0033] Step S2.3, based on low-level feature maps The SoftMax function is used to project it onto the probability space to obtain the corresponding probability distribution tensor. Furthermore, based on the definition of mutual information, a network structure is designed to calculate the similarity tensor at the disparity probability distribution level. Furthermore, the mutual volume is calculated using 3D convolution, with the specific formula as follows:

[0034]

[0035] Step S2.4: Use Hadamard multiplication to aggregate texture volume, correlation volume, and cross volume to calculate the three-fold fusion cost volume. .

[0036] Preferably, in step S2.4, the resolution of the texture volume, correlation volume, mutual volume, and tri-fusion cost volume is 1 / 4 of the resolution of the original binocular image.

[0037] Preferably, step S3 includes: using cascaded multi-layer 3D convolutions to fuse the three layers into a cost volume without changing 1 / 4 of the resolution. Step by step with the feature maps of the other three levels of the multi-level feature maps Weighted aggregation yields the aggregated cost volume. .

[0038] Preferably, step S4 includes the following sub-steps:

[0039] Step S4.1, for The disparity probability values ​​at each location are sorted, and only the two disparity values ​​with the highest probabilities are considered. The weighted summation yields the disparity prediction for the corresponding location, calculated using the following formula:

[0040]

[0041] in, To determine the cost volume after sorting, only the two with the highest probabilities are considered. That is, the corresponding probability value;

[0042] Step S4.2, for Using nearest neighbor interpolation, the original resolution is restored to obtain the complete cost volume. ,right All disparity values ​​at each location ( The weighted summation yields the disparity prediction for the corresponding location, calculated using the following formula:

[0043]

[0044] Preferably, step S5 includes the following sub-steps:

[0045] Step S5.1, based on the baseline distance of the stereo cameras ,focal length and in At any given time, the parallax value at a specific location The depth value at a specific location is calculated. , and thus generate Depth map at any given moment;

[0046] Step S5.2, in At the same time, the depth value at the same location was calculated using the same method. Then, based on the principle of similar triangles, if The camera's height above the ground at any given moment is This allows us to obtain the change in road surface undulation. The calculation formula is as follows:

[0047]

[0048] This invention also provides an automotive suspension control system, based on the above-mentioned real-time binocular stereo matching method based on three-fusion cost volume, comprising:

[0049] Binocular vision module: used to execute the above-mentioned real-time binocular stereo matching method based on three-fusion cost volume to obtain the change in road surface undulation; the range of the change in road surface undulation is divided into multiple intervals, each interval corresponding to a preset suspension posture, and the control system adjusts the suspension posture in real time according to the actual situation.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. This invention proposes a lightweight three-dimensional volume construction strategy. The volume constructed by this strategy includes multi-scale texture information, pixel-level correlation information, and probability distribution-level similarity information. It can effectively integrate information that is conducive to matching in the left and right views while maintaining low computational complexity of the model, thereby improving the model's ability to model three-dimensional features.

[0052] 2. This invention proposes a binocular stereo matching method based on three-dimensional cost volume fusion. Based on this method, disparity maps and depth maps can be accurately obtained, thereby providing accurate prior results for the input of road height during suspension control.

[0053] 3. This invention proposes a vehicle suspension control system based on road surface depth perception. This system can adaptively adjust the suspension posture based on changes in road surface undulation, which can significantly enhance the working stability and real-time performance of the vehicle suspension system. Attached Figure Description

[0054] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0055] Figure 1 This is a flowchart of a real-time binocular stereo matching method based on three-fusion cost volume in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the TCMNet model framework and its main components according to an embodiment of the present invention.

[0057] Figure 3 This is a schematic diagram of the suspension posture preset in an embodiment of the present invention;

[0058] Figure 4 This is a visualization of the matching results in an embodiment of the present invention;

[0059] Figure 5 This is a performance-computational complexity evaluation chart of the method of the present invention and six other methods in the embodiments of the present invention. Detailed Implementation

[0060] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0061] This invention provides a real-time binocular stereo matching method and a method-based automotive suspension control system. The aim is to extract multi-scale texture information, adaptively extract pixel-level correlations using a deformable attention mechanism, and further calculate probability distribution-level similarity to construct an efficient and information-rich three-fusion cost volume. This three-fusion cost volume maintains good stereo matching performance while keeping computationally low. Based on this, hierarchical control of the automotive suspension is further implemented to ensure driving comfort.

[0062] Example 1:

[0063] Figure 1 This is a flowchart of a real-time binocular stereo matching method based on three-fusion cost volume in an embodiment of the present invention.

[0064] like Figure 1 As shown, this embodiment provides a real-time binocular stereo matching method based on three-way fusion cost volume, including the following steps:

[0065] Step S1 involves processing the original binocular image to obtain multi-level feature maps in the form of three-dimensional tensors at different scales, thereby acquiring semantic and texture information at different levels.

[0066] Specifically, step S1 includes the following sub-steps:

[0067] Step S1.1: Regularize the original binocular image and expand it into a four-dimensional tensor to obtain the processed image.

[0068] In this embodiment, the regularization factors applied to the original binocular image are as follows: Where R, G, and B represent the three channels of the image, respectively. This represents the mean. Indicates variance.

[0069] Step S1.2: The processed image is passed through a lightweight feature extraction network, and initial multi-scale feature maps are obtained by progressively using convolution, pooling, and activation layers. .

[0070] Step S1.3, based on the initial multi-scale feature map, and using the structure of the feature pyramid, perform the following operations sequentially: ① For After upsampling and By splicing and combining, we obtain ; ② Further encoding using 1×1 convolutions yields new values ​​at the current level. ③ Thus, multi-level feature maps are obtained. .

[0071] In this embodiment, multi-level feature maps These correspond to feature information at resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32, respectively, and the size of their feature dimensions corresponds to the maximum disparity value that a binocular camera can capture. Relatedly, the feature map with 1 / 4 resolution is a low-level feature map, and the feature maps with the remaining resolution are feature maps of other levels.

[0072] Step S2: Based on the original binocular image and the low-level feature map of the multi-level feature map, construct a three-fusion cost volume to construct the similarity and correlation of the left and right views in the depth direction.

[0073] Specifically, step S2 includes the following sub-steps:

[0074] Step S2.1, based on the original stereo images A convolutional-regularized module with shared weights is used to extract texture information at the original resolution. Then, a cascaded, cross-layer connected convolutional neural network is designed to encode the texture information at the original resolution and extract the texture information at the low resolution. .

[0075] Will With low-level feature maps Aggregation, using 3D convolution to calculate texture volume The calculation formula is as follows:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081] in, and These are all intermediate outputs of the network. This indicates a normalization operation. Represents two-dimensional convolution. Represents 3D convolution. Represents the Hadamard product of matrices. This represents the tensor concatenation operation. This represents a compound operation of addition, convolution, and regularization.

[0082] Step S2.2, based on low-level feature maps First, use cosine similarity to roughly calculate the correlation tensor between the two. Then, using deformable attention mechanisms, for The location of each pixel in eigenvectors It adaptively searches for and matches other relevant pixels, and after weighted aggregation of the feature information of each relevant pixel, obtains a new feature vector. Repeat this operation for each position to obtain a new correlation tensor. The correlation volume is calculated using 3D convolution. The calculation formula is as follows:

[0083]

[0084] in, Linear Represents a linear layer. For correlation tensor The intermediate layer output obtained after further learning through two-dimensional convolution. This represents the bias in adaptive learning. This represents the aggregated weights after adaptive learning.

[0085] Step S2.3, based on low-level feature maps The SoftMax function is used to project it onto the probability space to obtain the corresponding probability distribution tensor. Furthermore, based on the definition of mutual information, a network structure is designed to calculate the similarity tensor at the disparity probability distribution level. Furthermore, the mutual volume is calculated using 3D convolution, with the specific formula as follows:

[0086]

[0087] Step S2.4: Use Hadamard multiplication to aggregate texture volume, correlation volume, and cross volume to calculate the three-fold fusion cost volume. .

[0088] In this embodiment, the resolutions of texture volume, correlation volume, mutual volume, and tri-fusion cost volume are all 1 / 4 of the resolution of the original binocular image.

[0089] Step S3: Based on the three-fusion cost volume and other layer feature maps of the multi-level feature maps, downsampling and upsampling are performed layer by layer according to the tensor size using 3D convolution to aggregate the cost volume and obtain the aggregated cost volume.

[0090] Specifically, step S3 includes: using cascaded multi-layer 3D convolutions to fuse the three layers into a cost volume without changing the 1 / 4 resolution. Step by step with the feature maps of the other three levels of the multi-level feature maps Weighted aggregation yields the aggregated cost volume. .

[0091] Step S4: Input the aggregated cost volume into the dual-scale disparity regression module, and use the arg-softmax function to output disparity maps at 1 / 4 resolution and original resolution, respectively.

[0092] Specifically, step S4 includes the following sub-steps:

[0093] Step S4.1, for The disparity probability values ​​at each location are sorted, and only the two disparity values ​​with the highest probabilities are considered. The weighted summation yields the disparity prediction for the corresponding location, calculated using the following formula:

[0094]

[0095] in, To determine the cost volume after sorting, only the two with the highest probabilities are considered. That is, the corresponding probability value;

[0096] Step S4.2, for Using nearest neighbor interpolation, the original resolution is restored to obtain the complete cost volume. .right All disparity values ​​at each location ( The weighted summation yields the disparity prediction for the corresponding location, calculated using the following formula:

[0097]

[0098] Step S5: Using the binocular camera parameters and disparity map, calculate the depth map of the road surface. By comparing the depth changes in the depth map in consecutive frames, calculate the change in road surface undulation based on geometric relationships, and use it as the input signal for subsequent suspension control.

[0099] Specifically, step S5 includes the following sub-steps:

[0100] Step S5.1, based on the baseline distance of the stereo cameras ,focal length and in At any given time, the parallax value at a specific location The depth value at a specific location is calculated. , and thus generate Depth map at any given moment.

[0101] Step S5.2, in At the same time, the depth value at the same location was calculated using the same method. Then, based on the principle of similar triangles, if The camera's height above the ground at any given moment is This allows us to obtain the change in road surface undulation. The calculation formula is as follows:

[0102]

[0103] Example 2:

[0104] This embodiment provides an automotive suspension control system, employing a real-time binocular stereo matching method based on three-way cost volume as described in Embodiment 1 above, including:

[0105] Binocular vision module: used in the real-time binocular stereo matching method based on three-fusion cost volume in the above embodiment 1 to obtain the change in road surface undulation; the range of the change in road surface undulation is divided into multiple intervals, each interval corresponding to a preset suspension posture, and the control system adjusts the suspension posture in real time according to the actual situation to ensure driving comfort.

[0106] Figure 3 This is a schematic diagram of the suspension posture preset in an embodiment of the present invention.

[0107] like Figure 3As shown, during actual vehicle operation, the changes in road surface undulation are monitored in real time. When the changes fall within a certain range, the suspension posture is immediately adjusted to the preset state of that range, so that the suspension can respond promptly to the dynamic changes in road conditions, ensuring the comfort and stability of vehicle operation.

[0108] This invention realizes a binocular stereo matching method with low matching error. The method demonstrated excellent matching performance in testing, with prediction results highly consistent with actual results, such as... Figure 4 As shown. This method not only maintains low endpoint error (EPE), first-order error (D1), and error beyond n pixels (>npx), but also has extremely low computational complexity. Test results and performance parameter comparisons are shown below. Figure 5 Given the crucial role of binocular stereo matching in environmental perception tasks and subsequent active suspension control safety, this invention has significant practical value, especially in improving the safety and efficiency of intelligent driving systems and enhancing vehicle ride comfort.

[0109] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0110] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A real-time binocular stereo matching method based on three-way fusion cost volume, characterized in that, Includes the following steps: Step S1: Based on the original binocular images, obtain multi-level feature maps in the form of three-dimensional tensors at different scales to obtain semantic and texture information at different levels; Step S1 includes the following sub-steps: Step S1.1: Regularize the original binocular image and expand it into a four-dimensional tensor to obtain the processed image; Step S1.2: The processed image is passed through a lightweight feature extraction network, and initial multi-scale feature maps are obtained by progressively using convolution, pooling, and activation layers. ; Step S1.3, based on the initial multi-scale feature map, and with the help of the feature pyramid structure, perform the following operations in sequence: ① For After upsampling and By splicing and combining, we obtain ; ② Further encoding using 1×1 convolutions yields new values ​​at the current level. ③ Thus, multi-level feature maps are obtained. ; Step S2: Based on the original binocular image and the low-level feature map of the multi-level feature map, construct a three-fusion cost volume to construct the similarity and correlation of the left and right views in the depth direction; Step S3: Based on the three-fusion cost volume and the other level feature maps of the multi-level feature map, downsampling and upsampling are performed layer by layer according to the tensor size using 3D convolution to aggregate the cost volume and obtain the aggregated cost volume. Step S4: Input the aggregated cost volume into the dual-scale disparity regression module, and use the arg-softmax function to output disparity maps at 1 / 4 resolution and original resolution respectively; Step S5: Using the binocular camera parameters and the disparity map, a depth map of the road surface is calculated. By comparing the depth changes in the depth map in consecutive frames, the change in road surface undulation is calculated based on geometric relationships and used as the input signal for subsequent suspension control.

2. The real-time binocular stereo matching method based on three-way fusion cost volume according to claim 1, characterized in that: In step S1.1, the regularization factors applied to the original binocular image are as follows: Where R, G, and B represent the three channels of the image, respectively. This represents the mean. Indicates variance.

3. The real-time binocular stereo matching method based on three-way fusion cost volume according to claim 2, characterized in that: In step S1.3, the multi-level feature map These correspond to feature information at resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32, respectively. Its feature dimension size is related to the maximum parallax value that a binocular camera can capture. Relatedly, the feature map with 1 / 4 resolution is a low-level feature map, and the feature maps with the remaining resolution are feature maps of other levels.

4. The real-time binocular stereo matching method based on three-way fusion cost volume as described in claim 1, characterized in that: Step S2 includes the following sub-steps: Step S2.1, based on the original stereo images The algorithm uses a shared-weight module to extract texture information at the original resolution, and then designs a cascaded, cross-layer connected convolutional neural network to encode the texture information at the original resolution, thereby extracting the texture information at the low resolution. ,Will With the low-level feature map Aggregation, using 3D convolution to calculate texture volume The calculation formula is as follows: in, and These are all intermediate outputs of the network. This indicates a normalization operation. Represents two-dimensional convolution. Represents 3D convolution. Represents the Hadamard product of matrices. This represents the tensor concatenation operation. This represents a compound operation of addition, convolution, and regularization. Step S2.2, based on the low-level feature map First, use cosine similarity to roughly calculate the correlation tensor between the two. Then, using a deformable attention mechanism, for The location of each pixel in eigenvectors It adaptively searches for and matches other relevant pixels, and after weighted aggregation of the feature information of each relevant pixel, obtains a new feature vector. Repeat this operation for each position to obtain a new correlation tensor. The correlation volume is calculated using 3D convolution. The calculation formula is as follows: in, Linear Represents a linear layer. For correlation tensor The intermediate layer output obtained after further learning through two-dimensional convolution. This represents the bias in adaptive learning. This represents the aggregated weights after adaptive learning; Step S2.3, based on the low-level feature map The SoftMax function is used to project it onto the probability space to obtain the corresponding probability distribution tensor. Furthermore, based on the definition of mutual information, a network structure is designed to calculate the similarity tensor at the disparity probability distribution level. Furthermore, the mutual volume is calculated using 3D convolution, with the specific formula as follows: Step S2.4: Use Hadamard multiplication to aggregate the texture volume, the correlation volume, and the cross volume to calculate the three-fold fusion cost volume. .

5. The real-time binocular stereo matching method based on three-way fusion cost volume as described in claim 4, characterized in that: In step S2.4, the resolutions of the texture volume, the correlation volume, the mutual volume, and the tri-fusion cost volume are all 1 / 4 of the resolution of the original binocular image.

6. The real-time binocular stereo matching method based on three-way fusion cost volume according to claim 1, characterized in that: Step S3 includes: using cascaded multi-layer 3D convolutions to fuse the three layers into a cost volume without changing the 1 / 4 resolution. Step by step, the feature maps of the other three levels of the multi-level feature map are compared. Weighted aggregation yields the aggregated cost volume. .

7. The real-time binocular stereo matching method based on three-way fusion cost volume according to claim 1, characterized in that: Step S4 includes the following sub-steps: Step S4.1, for The disparity probability values ​​at each location are sorted, and only the two disparity values ​​with the highest probabilities are considered. The weighted summation yields the disparity prediction for the corresponding location, calculated using the following formula: in, To determine the cost volume after sorting, only the two with the highest probabilities are considered. That is, the corresponding probability value; Step S4.2, for Using nearest neighbor interpolation, the original resolution is restored to obtain the complete cost volume. ,right All disparity values ​​at each location ( The weighted summation yields the disparity prediction for the corresponding location, calculated using the following formula: 。 8. The real-time binocular stereo matching method based on three-way fusion cost volume according to claim 1, characterized in that: Step S5 includes the following sub-steps: Step S5.1, based on the baseline distance of the stereo cameras ,focal length and in At any given time, the parallax value at a specific location The depth value at the specific location is calculated. , and thus generate Depth map at any given moment; Step S5.2, in At the same time, the depth value at the same location was calculated using the same method. Then, based on the principle of similar triangles, if The camera's height above the ground at any given moment is This allows us to obtain the change in road surface undulation. The calculation formula is as follows: 。 9. A vehicle suspension control system, based on the real-time binocular stereo matching method based on three-way cost volume as described in any one of claims 1 to 8, characterized in that, include: Binocular vision module: used to execute the real-time binocular stereo matching method based on three-fusion cost volume as described in any one of claims 1 to 8, to obtain the change in road surface undulation; to divide the range of the change in road surface undulation into multiple intervals, each interval corresponding to a preset suspension posture, and to adjust the suspension posture in real time according to the actual situation.

Images