Electric pleasure boat target detection and distance measurement system based on binocular vision

Abstract

The invention discloses an electric pleasure boat target detection and distance measurement system based on binocular vision. The system comprises a binocular camera shooting assembly; according to the invention, through the algorithm optimization module and the shared convolution module, the network structure is optimized, the fusion capability of multi-scale features is enhanced, the model parameter quantity is greatly reduced, and the calculation efficiency and precision are improved; the importance of each channel is adaptively adjusted through the channel attention module and the space attention module, so that the model can better pay attention to an important area in an image and better capture the relationship between different channels, and the expression ability and the sensitivity of features are improved; by using a novel global channel space attention module, a global dependency relationship in a feature map can be comprehensively captured, so that the performance of the model is further improved; through the map storage assembly, the route of the water area can be recorded and updated, the navigation route can be rapidly provided for the water area without large changes during repeated navigation, and the use experience of a user is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection and ranging of electric cruise ships, and in particular to a target detection and ranging system for electric cruise ships based on binocular vision. Background Art

[0002] An electric cruise ship is a water vehicle that is powered by electricity and is widely used for sightseeing in lakes, rivers, reservoirs and other waters. In water parks, scenic spots, parks and other places, electric cruise ships have also become an important choice for tourists' leisure and entertainment. Some electric cruise ships also have catering, accommodation and other functions to provide tourists with more comprehensive services. The target detection and ranging technology applied to electric cruise ships can monitor the environment around the cruise ship in real time, including other cruise ships, water obstacles (such as reefs, bridges, etc.) and changes in the water surface. Through timely detection and ranging, the system can issue a collision warning in advance to avoid collisions between cruise ships and other objects, thereby ensuring the safety of tourists and cruise ships. However, the traditional electric cruise ship target detection and ranging system still has shortcomings. First, the water navigation environment is complex, and small target obstacles (such as buoys and floating objects) are difficult to detect, mainly because the target First, factors such as small landmarks, background interference, and illumination changes affect the accuracy of existing algorithms; second, in existing algorithms, the ASPP module captures multi-scale information through dilated convolutions with different dilation rates, but when the dilation rate is too large, the effective weights in the convolution kernel will decrease. For example, a 3×3 convolution kernel may only have a small part in the center actually involved in the calculation, while the weights on the edges are invalid due to the dilation (i.e., zero padding), which will cause the network to be unable to effectively capture global contextual information, and thus unable to fully utilize all available convolution kernel parameters, thereby limiting the expansion effect of its receptive field and its feature expression capabilities; third, it is difficult for existing algorithms to fully capture the global dependencies in feature maps, making it difficult for the GCSA module to extract and utilize key information in feature maps, thus affecting the performance of the model; therefore, it is necessary to design an electric cruise ship target detection and ranging system based on binocular vision. Summary of the invention

[0003] The purpose of the present invention is to provide an electric cruise ship target detection and ranging system based on binocular vision to solve the problems raised in the above background technology.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: an electric cruise ship target detection and ranging system based on binocular vision, comprising a binocular camera assembly, wherein the binocular camera assembly is composed of an image acquisition module, an image processing module, a noise processing module, a contrast enhancement module and an image output module, wherein the image acquisition module is control-connected to the image processing module, the image processing module is control-connected to the noise processing module, the noise processing module is control-connected to the image output module, the image output module is control-connected to the contrast enhancement module, and the contrast enhancement module is control-connected to the noise processing module.

[0005] As a further technical solution of the present invention, the image output module is controlled to be connected to an image receiving module in the detection component.

[0006] As a further technical solution of the present invention, the detection component is composed of an image receiving module, a feature processing module, an algorithm optimization module, a shared convolution module, a feature path module, a boundary regression module, a category prediction module, a scaling adjustment module and a feature output module. The image receiving module is controlled to be connected with the feature processing module, the feature processing module is controlled to be connected with the algorithm optimization module, the algorithm optimization module is controlled to be connected with the feature path module, the feature path module is controlled to be connected with the shared convolution module, the shared convolution module is controlled to be connected with the feature processing module, the feature path module is controlled to be connected with the boundary regression module, the boundary regression module is controlled to be connected with the scaling adjustment module, the feature path module is controlled to be connected with the category prediction module, the category prediction module is controlled to be connected with the feature output module, and the feature output module is controlled to be connected with the boundary regression module.

[0007] As a further technical solution of the present invention, the feature output module controls the classification receiving module connected to the classification component.

[0008] As a further technical solution of the present invention, the classification component is composed of a classification receiving module, a segmentation processing module, an architecture optimization module, an ASPP module, a GCSA module and a classification output module. The classification receiving module is controlled to be connected with the segmentation processing module, the segmentation processing module is controlled to be connected with the architecture optimization module, the architecture optimization module is controlled to be connected with the classification output module, the classification output module is controlled to be connected with the ASPP module, the ASPP module is controlled to be connected with the segmentation processing module, the segmentation processing module is controlled to be connected with the GCSA module, and the GCSA module is controlled to be connected with the classification output module.

[0009] As a further technical solution of the present invention, the classification output module controls the feature receiving module in the matching refining component.

[0010] As a further technical solution of the present invention, the matching and refining component is composed of a feature receiving module, a feature extraction module, a feature description module, a feature matching module, an abnormality removal module, a fitting and refining module and a refining output module. The feature receiving module is controlled to be connected with a feature extraction module, the feature extraction module is controlled to be connected with a feature description module, the feature description module is controlled to be connected with a feature matching module, the feature matching module is controlled to be connected with an abnormality removal module, the abnormality removal module is controlled to be connected with a fitting and refining module, and the fitting and refining module is controlled to be connected with a refining output module.

[0011] As a further technical solution of the present invention, the refined output module controls the calculation receiving module in the distance calculation component.

[0012] As a further technical solution of the present invention, the distance calculation component is composed of a calculation receiving module, a calculation processing module and a distance output module. The calculation receiving module is controlled to be connected with the calculation processing module, the calculation processing module is controlled to be connected with the distance output module, the distance output module is controlled to be connected with the map update module, the map update module is controlled to be connected with the multi-threaded module, and the map update module is controlled to be connected with the map receiving module in the map storage component.

[0013] As a further technical solution of the present invention, the map update module is controlled to be connected with a multi-threaded module map storage component which is composed of a map receiving module, a map comparison module, a route update module and a route storage module. The map receiving module is controlled to be connected with a map comparison module, the map comparison module is controlled to be connected with a route update module, and the route update module is controlled to be connected with a route storage module.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: the electric cruise ship target detection and ranging system based on binocular vision optimizes the network structure, reduces the number of parameters and model size, and enhances the fusion capability of multi-scale features through an algorithm optimization module and a shared convolution module, and greatly reduces the number of model parameters by using the shared convolution technology, thereby improving the computational efficiency and accuracy; the importance of each channel can be adaptively adjusted through the channel attention module and the spatial attention module in the CBAM attention mechanism, so that the model can better focus on important areas in the image and better capture the relationship between different channels, which will give higher weights to the features of key channels when ASPP extracts multi-scale features, thereby improving the expressiveness and receptive field of the features; by using a novel global channel spatial attention module GCSA, the global dependencies in the feature map can be fully captured, so that the GCSA module can effectively extract and utilize the key information in the feature map, thereby further improving the performance of the model, receiving map information through the map storage component, being able to record and update the route of the water area, and being able to quickly provide the navigation route for the water area without significant changes during repeated navigation, thereby improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the overall architecture diagram of the present invention;

[0016] Figure 2 This is a flow chart of the binocular camera assembly in the present invention;

[0017] Figure 3 It is a flow chart of the detection component in the present invention;

[0018] Figure 4 A flow chart of the classification components in the present invention;

[0019] Figure 5A flow chart of the matching and refining components in the present invention;

[0020] Figure 6 is a flow chart of the distance calculation component in the present invention;

[0021] Figure 7 A flowchart of a map storage component in the present invention;

[0022] Figure 8 It is a network structure diagram of C2f-SBS in the present invention;

[0023] Fig. 9 It is the network structure diagram of SCLD in the present invention;

[0024] Fig.10 It is the network structure diagram of MobilenetV2 in the present invention;

[0025] Fig.11 This is the network structure diagram of CBAM-ASPP in the present invention;

[0026] Fig.12 It is the GCSA global channel space attention structure diagram in the present invention;

[0027] Fig.13 is a reference diagram for distance measurement in the present invention;

[0028] In the figure: 1. Binocular camera component; 2. Detection component; 3. Classification component; 4. Matching and refining component; 5. Distance calculation component; 6. Map update module; 7. Multithreading module; 8. Map storage component; 11. Image acquisition module; 12. Image processing module; 13. Noise processing module; 14. Contrast enhancement module; 15. Image output module; 21. Image receiving module; 22. Feature processing module; 23. Algorithm optimization module; 24. Shared convolution module; 25. Feature path module; 26. Boundary regression module; 27. Category prediction module; 28. Scaling adjustment module; 29. ​​Feature Output module; 31. Classification receiving module; 32. Segmentation processing module; 33. Architecture optimization module; 34. ASPP module; 35. GCSA module; 36. Classification output module; 41. Feature receiving module; 42. Feature extraction module; 43. Feature description module; 44. Feature matching module; 45. Abnormal elimination module; 46. Fitting and refining module; 47. Refining output module; 51. Calculation receiving module; 52. Calculation processing module; 53. Distance output module; 81. Map receiving module; 82. Map comparison module; 83. Route update module; 84. Route storage module. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] Please see attached Figure 1 -Attached Fig.13The present invention provides an embodiment: an electric cruise ship target detection and ranging system based on binocular vision, comprising a binocular camera assembly 1, the binocular camera assembly 1 is composed of an image acquisition module 11, an image processing module 12, a noise processing module 13, a contrast enhancement module 14 and an image output module 15, the image acquisition module 11 is controlled to be connected to the image processing module 12, the image processing module 12 is controlled to be connected to the noise processing module 13, the noise processing module 13 is controlled to be connected to the image output module 15, the image output module 15 is controlled to be connected to the contrast enhancement module 14, the contrast enhancement module 14 is controlled to be connected to the noise processing module 13, the binocular camera assembly 1 is respectively arranged in two waterproof cameras, and the cameras are connected to the It is installed in front of the top of the ship, and the left and right images are collected by the image acquisition module 11 and transmitted to the image processing module 12. The image processing module 12 distributes the collected images to the noise processing module 13 and the contrast enhancement module 14 for preprocessing. The mean filtering method is used in the noise processing module 13 to smooth the image and remove noise by calculating the statistical characteristics of the average value of the neighborhood pixels. The histogram equalization method is used in the contrast enhancement module 14 to adjust the grayscale distribution in the image so that the grayscale value of the image is more evenly distributed, thereby improving the contrast and making the target and background in the image clearer and more distinguishable. Then the image after denoising and contrast enhancement is transmitted to the image output module 15 for output. The image output module 15 is controlled and connected with the image receiving module 21 in the detection component 2, and the detection component 2 is composed of the image receiving module 21, the feature processing module 22, the algorithm optimization module 23, the shared convolution module 24, the feature path module 25, the boundary regression module 26, the category prediction module 27, the scaling adjustment module 28 and the feature output module 29. The image receiving module 21 is controlled and connected with the feature processing module 22, the feature processing module 22 is controlled and connected with the algorithm optimization module 23, the algorithm optimization module 23 is controlled and connected with the feature path module 25, the feature path module 25 is controlled and connected with the shared convolution module 24, the shared convolution module 24 is controlled and connected with the feature processing module 22, and the feature path module 25 is controlled and connected with the boundary regression module 26, the category prediction module 27, the scaling adjustment module 28 and the feature output module 29. The boundary regression module 26 is controlled to be connected to the scaling adjustment module 28. The feature path module 25 is controlled to be connected to the category prediction module 27. The category prediction module 27 is controlled to be connected to the feature output module 29. The feature output module 29 is controlled to be connected to the boundary regression module 26. The image receiving module 21 is used to receive the processed image output by the image output module 15 and transmit it to the feature processing module 22. The feature processing module 22 distributes the image to the algorithm optimization module 23 and the shared convolution module 24 for processing. In the algorithm optimization module 23, the StarBlocks module in the StarNet network architecture is first introduced. The StarBlocks output form is:

[0031] x=f DWC (F)

[0032] y=x+f C (ReLU(f C (x))f C (x)

[0033] Among them, F represents the input feature, x represents the result of preliminary feature processing, and f DWC represents the depthwise separable convolution function, y is the output of StarBlocks, and f C Represents the operation of convolution plus batch normalization (BN), and ReLU represents the activation function;

[0034] Based on StarNet, a C2f-SBS module is proposed. By replacing the traditional Bottleneck structure with StarBlocks, the star operation of StarNet is integrated to efficiently extract high-dimensional and nonlinear features in low-dimensional space. In the shared convolution module 24, a shared convolution lightweight detection head SCLD is proposed. By receiving the feature maps of different stages and scales of the algorithm optimization module 23, numbered P3, P4 and P5, a 1×1 convolution is first applied to each feature map, combined with a group normalization GN operation, and the stride of the 1×1 convolution is set to 1 to maintain the spatial dimension of the feature map. After the initial 1×1 convolution, the feature map is sent to the shared convolution layer, which is composed of two 3×3 convolution layers. The first Conv-GN layer extracts key spatial features while maintaining the receptive field, and then passes through two consecutive 3×3 Conv-GN layers to further enhance the feature expression ability. The processed feature map is divided into two paths through the feature path module 25. One path is transmitted to the boundary regression module 26, and the specific position of the target is estimated through the boundary regression module 26. The other path is transmitted to the category prediction module 27, and the classification task is processed by the category prediction module 27. A learnable scaling layer is designed in the scaling adjustment module 28. The scaling layer adjusts the scale of the output feature map through element-by-element operations, and acts in the boundary regression module 26 to stabilize the edge. The bounding box is predicted, and finally the image processed by the target detection is output through the feature output module 29. The feature output module 29 controls the connection with the classification receiving module 31 in the classification component 3. The classification component 3 is composed of the classification receiving module 31, the segmentation processing module 32, the architecture optimization module 33, the ASPP module 34, the GCSA module 35 and the classification output module 36. The classification receiving module 31 controls the connection with the segmentation processing module 32, the segmentation processing module 32 controls the connection with the architecture optimization module 33, the architecture optimization module 33 controls the connection with the classification output module 36, the classification output module 36 controls the connection with the ASPP module 34, the ASPP module 34 controls the connection with the segmentation processing module 32, and the segmentation processing module 32 controls the connection with the classification output module 36. It is connected to a GCSA module 35, which controls the connection to a classification output module 36. The classification receiving module 31 receives the target detected image output by the feature output module 29 and transmits it to the segmentation processing module 32. The segmentation processing module 32 uses an improved semantic segmentation algorithm to perform pixel-level classification on the image. In the process, MobilenetV2 is used in the architecture optimization module 33 to replace Xception as the backbone network of the improved Deeplabv3+. The deep separable convolution technology composed of deep convolution and point-by-point convolution in MobileNetV1 is used. The number of parameters and the computational complexity are significantly lower than those of Xception. An inverted residual structure is used, which is different from the traditional residual structure.First, the number of channels of the feature map is increased by 1×1 convolution to enhance the network processing capability, and then the important features in the image are extracted by 3×3 convolution. Finally, the dimensionality is reduced to restore the original pixels of the image by 1×1 convolution. This inverted residual structure helps to maintain the integrity of the features and improve the expressiveness of the model. In the last layer of the inverted residual structure, MobileNetV2 uses a linear activation function such as the linear part of ReLU6 to maintain the linear separability of the features and avoid information loss. In the ASPP module 34, the CBAM attention mechanism is introduced in the atrous spatial convolution pooling pyramid (ASPP). The channel attention module and the spatial attention module in the CBAM attention mechanism can adaptively adjust the importance of each channel, so that the model can better focus on the important areas in the image and better capture the relationship between different channels. This will give higher weights to the features of the key channels when ASPP extracts multi-scale features, thereby improving the expressiveness and receptive field of the features. The CBAM-ASPP network structure first uses the ASPP module 34 to perform multi-scale spatial convolution operations on the input features to generate feature maps of different scales. Subsequently, these feature maps are sent to the CBA The channel attention weighted processing is performed in the M module, and finally the weighted feature representation is obtained. The GCSA module 35 is applied to the pooling layer in the ASPP. In the GCSA module 35, the input feature map is first sent to the channel attention submodule and then to the spatial attention submodule. The initial feature map contains multiple channels, and the spatial size of each channel is H×W. In the channel attention submodule, the input feature map is first dimensionalized, and it is converted from C×H×W (number of channels×height×width) to W×H×C for subsequent processing; secondly, a multilayer perceptron (M) consisting of two layers is used. LP) is used to capture the dependencies between channels. The first layer of MLP reduces the number of channels to one-fourth of the original number. Then, the nonlinear characteristics are introduced through the ReLU activation function. Then, the second layer of MLP restores the number of channels to the original dimension to achieve effective integration of information. After completing the capture of the dependencies between channels, the inverse dimensionality transformation is performed to restore the feature map to the form of C×H×W. Finally, the channel attention map is generated through the Sigmoid activation function, which reflects the importance of different channels. The input feature map is multiplied element by element with the channel attention map to obtain the enhanced feature map. The process is as follows:

[0035] F channel =σ(MLP(Permute(F input )))⊙F input Among them, F channel is the enhanced feature map, σ is the Sigmoid function, ⊙ is the element-by-element multiplication, F input is the original input feature map;

[0036] The channel shuffle operation is introduced to further enhance the mixing and sharing of feature information. First, the enhanced feature map is evenly divided into 4 groups along the channel dimension, each group contains ( C / 4 ) channels, and then perform a transposition operation on the feature maps in each group to disrupt the channel order, thereby realizing information exchange and reorganization between channels, and finally reassemble the disrupted feature maps and restore them to the original ( C×H×W ) Shape, this channel shuffling strategy helps promote information interaction between different channels, thereby improving the expressiveness of features and the generalization performance of the model. The specific process is:

[0037] F shuffle =ChannelShuffle(F channel )

[0038] Among them, F shuffle is the shuffled feature map, F channel is the number of channels of the input feature map;

[0039] In the spatial attention submodule, the input feature map first passes through a 7×7 convolution layer (Conv layer), which reduces the number of channels to 1 / 4 of the original, thereby achieving feature dimensionality reduction. Subsequently, the features are normalized through a batch normalization layer (BN layer) to eliminate internal covariate shifts and improve the training stability of the model. Next, the ReLU activation function is used for nonlinear transformation to increase the expressive power of the model. Then, the feature map passes through a second 7×7 convolution layer to restore the number of channels to the original dimension C to achieve feature dimensionality increase. The features are normalized again through a batch normalization layer. Finally, the sigmoid activation function is used to generate a spatial attention map, which can reflect the importance of features at different spatial positions. The shuffled feature map is multiplied element-by-element with the spatial attention map to obtain the final output feature map that incorporates the spatial attention information. The specific process is as follows:

[0040] F spatial =σ(Conv(BN)(ReLU(Conv(F shuffle ))))⊙F shuffle

[0041] Among them, F spatial It is the feature map after spatial attention processing;

[0042] The final output feature map includes enhanced features after channel attention, channel shuffling and spatial attention, and is output through the classification output module 36. The classification output module 36 controls the connection with the feature receiving module 41 in the matching and refining component 4. The matching and refining component 4 is composed of a feature receiving module 41, a feature extraction module 42, a feature description module 43, a feature matching module 44, an abnormality rejection module 45, a fitting and refining module 46 and a refining output module 47. The feature receiving module 41 controls the connection with the feature extraction module 42, the feature extraction module 42 controls the connection with the feature description module 43, the feature description module 43 controls the connection with the feature matching module 44, the feature matching module 44 controls the connection with the abnormality rejection module 45, and the abnormality rejection module The block 45 is connected to the fitting and refining module 46, and the fitting and refining module 46 is connected to the refining output module 47. The feature receiving module 41 receives the feature map output by the classification output module 36, extracts feature points by using the ORB feature extraction algorithm in the feature extraction module 42, and describes each feature point by the feature description module 43 to generate a feature descriptor for representing the local image information of the feature point. Then, the feature matching module 44 uses the feature descriptor to match the left and right images to find the corresponding feature point pairs, and then uses the RANSAC algorithm through the abnormal rejection module 45 to randomly select a set of feature point pairs as model parameters, and then calculate the consistency of other feature point pairs with the model, so as to screen out the internal points (correct matching). Points) and external points (abnormal matching points) are selected to remove abnormal matching points. After being processed by the RANSAC algorithm, a group of correct feature point pairs can be obtained for subsequent depth calculation. They are transmitted to the fitting and refining module 46. A two-dimensional quadratic surface fitting method is used. Based on the local image information around the feature points, a quadratic surface is fitted to approximate the real position of the feature points. The matched feature points are refined at the sub-pixel level. The refined feature points are output through the refining output module 47. The refining output module 47 controls the calculation receiving module 51 in the distance calculation component 5. The distance calculation component 5 is composed of the calculation receiving module 51, the calculation processing module 52 and the distance output module 53. The calculation receiving module 51 controls The control system is connected to a calculation processing module 52, and the calculation processing module 52 is controlled to be connected to a distance output module 53. The calculation receiving module 51 receives the feature point information output by the refined output module 47 and transmits it to the calculation processing module 52. In the calculation processing module 52, the distances of the three sides of the triangle formed by the left and right cameras and the target ship are set to A, B, and C respectively, and the angles formed by the three sides are α, β, and γ respectively. The midpoint position of the binocular camera system is regarded as the position of the ship equipped with the binocular camera, and the straight-line distance L between the target ship and the midpoint is used to represent the actual distance between the two ships. η is the angle between L and the positive direction of the x-axis of the camera plane. The known condition is the horizontal distance C of the camera, and the parameter to be determined is the ideal vertical distance H from the target ship to the camera screen. It is obtained by trigonometric function:

[0043] H=Asinβ=Bsinγ

[0044] From the triangle sine theorem we get:

[0045]

[0046] It follows that:

[0047]

[0048] Where η is the relative angle between the target and the ship;

[0049] To calculate H, we need to get the values ​​of α, β, and γ. β and γ are related to the camera's field of view. The field of view defines the width of the scene that the camera lens can capture. This feature is affected by the focal length of the lens, the size of the sensor, and the shooting distance. The shorter the focal length, the wider the field of view, and the more elements are included in the picture. The imaging plane is the specific mapping of this visual range on the sensor, satisfying the following relationship:

[0050] δ+2ε=π

[0051] Among them, δ is the horizontal field of view of the camera; we get ε l +θ l =β,ε r +θ r =γ, d1+d2+d3 are the horizontal pixels of the imaging plane corresponding to the left camera, d2+d3+d4 are the horizontal pixels of the imaging plane corresponding to the right camera, and d2 is the angle θ corresponding to the distance between the target ship and the left edge of the right imaging plane r d3 is the angle θ corresponding to the distance between the target ship and the right edge of the left imaging plane. l Pixels:

[0052]

[0053] C is the baseline between the two cameras, δ and ε are obtained by querying the camera parameters, d can be obtained by program calculation, and the focal length of the camera does not need to be obtained. The distance between the object and the camera plane can be calculated by this formula; and output to the map update module 6 through the distance output module 53, and a map is established through the map update module 6. The positions of the own ship and the identified ships are updated on the map according to the own ship positioning, the target ship distance and the angle. The distance output module 53 is controlled to be connected to the map update module 6, and the map update module 6 is controlled to be connected to the multi-thread module 7. The multi-thread module 7 adopts multi-thread technology to realize parallel processing of real-time target detection and depth calculation. The map update module 6 is controlled to be connected to the map receiving module 81 in the map storage component 8. The map storage component 8 is composed of a map receiving module 81, a map comparison module 82, a route update module 83 and a route storage module 84. The map receiving module 81 is controlled to be connected to the map comparison module 82, the map comparison module 82 is controlled to be connected to the route updating module 83, the route updating module 83 is controlled to be connected to the route storage module 84, the map information transmitted by the map updating module 6 is received by the map receiving module 81, and compared by the map comparison module 82, the navigation routes of the historical navigation waters are stored in the route storage module 84, when the map comparison module 82 determines that the current waters are not navigated or there are major changes in the historical navigation routes, the navigation routes of the current waters are recorded or updated by the route updating module 83, and stored by the route storage module 84, if the historical navigation routes of the current waters are stored in the route storage module 84, and the map comparison module 82 determines that the waters will not affect the historical navigation routes, the navigation routes are timely output by the route storage module 84.

[0054] Working principle: When in use, the algorithm optimization module 23 and the shared convolution module 24 are used to optimize the network structure, reduce the number of parameters and model size, and enhance the fusion capability of multi-scale features. The shared convolution technology is used to significantly reduce the number of model parameters and improve computational efficiency and accuracy. The channel attention module and the spatial attention module in the CBAM attention mechanism can adaptively adjust the importance of each channel, so that the model can better focus on important areas in the image and better capture the relationship between different channels. This will give higher weights to the features of key channels when ASPP extracts multi-scale features, thereby improving the expressiveness and receptive field of the features. By using a novel global channel spatial attention module GCSA, the global dependencies in the feature map can be fully captured, so that the GCSA module 35 can effectively extract and utilize the key features in the feature map. information, thereby further improving the performance of the model; the map information transmitted by the map update module 6 is received through the map receiving module 81, and compared through the map comparison module 82, and the navigation route of the historical navigation waters is stored in the route storage module 84. When the map comparison module 82 determines that the current waters are not navigated or there are major changes in the historical navigation routes, the navigation route of the current waters is recorded or updated through the route update module 83, and stored through the route storage module 84. If the historical navigation route of the current waters is stored in the route storage module 84, and the map comparison module 82 determines that the waters will not affect the historical navigation route, the navigation route is output in time through the route storage module 84, so that the route of the waters can be recorded and updated, and when the navigation is repeated, the navigation route can be quickly provided for the waters without major changes, thereby improving the user experience.

[0055] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An electric cruise ship target detection and ranging system based on binocular vision, comprising a binocular camera assembly (1), characterized in that: The binocular camera assembly (1) is composed of an image acquisition module (11), an image processing module (12), a noise processing module (13), a contrast enhancement module (14) and an image output module (15); the image acquisition module (11) is controllably connected to the image processing module (12); the image processing module (12) is controllably connected to the noise processing module (13); the noise processing module (13) is controllably connected to the image output module (15); the image output module (15) is controllably connected to the contrast enhancement module (14); and the contrast enhancement module (14) is controllably connected to the noise processing module (13).

2. The electric cruise ship target detection and ranging system based on binocular vision according to claim 1 is characterized in that: The image output module (15) controls and is connected to an image receiving module (21) in the detection component (2).

3. The electric cruise ship target detection and ranging system based on binocular vision according to claim 2 is characterized in that: The detection component (2) is composed of an image receiving module (21), a feature processing module (22), an algorithm optimization module (23), a shared convolution module (24), a feature path module (25), a boundary regression module (26), a category prediction module (27), a scaling adjustment module (28) and a feature output module (29), wherein the image receiving module (21) controls the connection with the feature processing module (22), the feature processing module (22) controls the connection with the algorithm optimization module (23), and the algorithm optimization module (23) controls the connection with the feature path module (25). ), the feature path module (25) is controlled to be connected to the shared convolution module (24), the shared convolution module (24) is controlled to be connected to the feature processing module (22), the feature path module (25) is controlled to be connected to the boundary regression module (26), the boundary regression module (26) is controlled to be connected to the scaling adjustment module (28), the feature path module (25) is controlled to be connected to the category prediction module (27), the category prediction module (27) is controlled to be connected to the feature output module (29), and the feature output module (29) is controlled to be connected to the boundary regression module (26).

4. The electric cruise ship target detection and ranging system based on binocular vision according to claim 3 is characterized in that: The feature output module (29) controls a classification receiving module (31) connected to the classification component (3).

5. The electric cruise ship target detection and ranging system based on binocular vision according to claim 4 is characterized in that: The classification component (3) is composed of a classification receiving module (31), a segmentation processing module (32), an architecture optimization module (33), an ASPP module (34), a GCSA module (35) and a classification output module (36); the classification receiving module (31) is controlled to be connected to the segmentation processing module (32); the segmentation processing module (32) is controlled to be connected to the architecture optimization module (33); the architecture optimization module (33) is controlled to be connected to the classification output module (36); the classification output module (36) is controlled to be connected to the ASPP module (34); the ASPP module (34) is controlled to be connected to the segmentation processing module (32); the segmentation processing module (32) is controlled to be connected to the GCSA module (35); and the GCSA module (35) is controlled to be connected to the classification output module (36).

6. The electric cruise ship target detection and ranging system based on binocular vision according to claim 5 is characterized in that: The classification output module (36) controls the feature receiving module (41) in the matching refining component (4) to be connected.

7. The electric cruise ship target detection and ranging system based on binocular vision according to claim 6 is characterized in that: The matching and refining component (4) is composed of a feature receiving module (41), a feature extraction module (42), a feature description module (43), a feature matching module (44), an abnormality elimination module (45), a fitting and refining module (46) and a refining output module (47). The feature receiving module (41) controls the connection with the feature extraction module (42), the feature extraction module (42) controls the connection with the feature description module (43), the feature description module (43) controls the connection with the feature matching module (44), the feature matching module (44) controls the connection with the abnormality elimination module (45), the abnormality elimination module (45) controls the connection with the fitting and refining module (46), and the fitting and refining module (46) controls the connection with the refining output module (47).

8. The electric cruise ship target detection and ranging system based on binocular vision according to claim 7 is characterized in that: The refined output module (47) controls the calculation receiving module (51) in the distance calculation component (5) connected thereto.

9. The electric cruise ship target detection and ranging system based on binocular vision according to claim 8 is characterized in that: The distance calculation component (5) is composed of a calculation receiving module (51), a calculation processing module (52) and a distance output module (53); the calculation receiving module (51) controls the connection with the calculation processing module (52); the calculation processing module (52) controls the connection with the distance output module (53); the distance output module (53) controls the connection with the map update module (6); the map update module (6) controls the connection with the multithreading module (7); and the map update module (6) controls the connection with the map receiving module (81) in the map storage component (8).

10. The electric cruise ship target detection and ranging system based on binocular vision according to claim 9, characterized in that: The map storage component (8) is composed of a map receiving module (81), a map comparison module (82), a route updating module (83) and a route storage module (84); the map receiving module (81) is controlled to be connected to the map comparison module (82); the map comparison module (82) is controlled to be connected to the route updating module (83); and the route updating module (83) is controlled to be connected to the route storage module (84).

Images