Point cloud target detection method based on laser radar fused sediment information

Abstract

The invention provides a point cloud target detection method based on laser radar fused sediment information, and relates to the technical field of laser radar point cloud data processing, and the method specifically comprises the following steps: enabling a sediment classification result to correspond to each point in a point cloud, enabling each point to belong to a class result, and obtaining a point cloud set with classification; and carrying out farthest point sampling on the point clouds to be classified to obtain key points. And taking the key point as a sphere center, setting the radius of the spherical region and the maximum number of sampling points in the spherical region, and carrying out sphere query operation to obtain a local point set. And inputting the local point set into a joint recognition network for target recognition, wherein the joint recognition network comprises a point cloud feature extraction module, a three-dimensional convolution module and an RPN network which are connected in sequence. And setting a total loss function of the joint recognition network, and evaluating a target recognition result. According to the technical scheme of the invention, the problems of weak generalization ability and weak robustness of a point cloud target detection method in an outdoor large scene in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the field of laser radar point cloud data processing, and in particular to a point cloud target detection method based on laser radar fused with bottom information. Background Art

[0002] The data set collected by the LiDAR is fused to obtain point cloud data. Point cloud data is a collection of sample points on the surface of the target under test obtained in the same spatial coordinate system. These points are represented by three-dimensional coordinates. The point cloud can display the surface feature information of the target under test. At the same time, depending on the measurement principle, the point cloud can also contain information such as laser reflection intensity and color. As for the target recognition method in point cloud data, with the rapid development of deep learning in recent years, point cloud target recognition using deep learning methods has gradually replaced point cloud target recognition based on traditional methods.

[0003] Although many methods of point cloud target recognition based on deep learning have emerged in recent years, most of them are based on point cloud target recognition in autonomous driving scenarios. For target recognition in outdoor large scene point clouds, there are often false detections and missed detections of targets. This is because outdoor large scenes often contain more complex environmental factors, such as terrain changes, abundant natural objects, etc. Therefore, point cloud target recognition in outdoor large scenes requires algorithms with better generalization ability and stronger robustness.

[0004] Therefore, there is a need for a point cloud target detection method based on lidar and fused with bottom information that is suitable for large outdoor scenes, has strong generalization ability and strong robustness. Summary of the invention

[0005] The main purpose of the present invention is to provide a point cloud target detection method based on laser radar fusion bottom information, so as to solve the problems of weak generalization ability and weak robustness of the point cloud target detection method in the prior art in large outdoor scenes.

[0006] To achieve the above object, the present invention provides a point cloud target detection method based on laser radar fusion bottom information, which specifically includes the following steps: S1, correspond the bottom classification result to each point in the point cloud, so that each point belongs to a category result, and obtain a point cloud set with classification.

[0007] S2, sampling the farthest point of the classified point cloud obtained in step S1 to obtain key points.

[0008] S3, taking the key point as the sphere center, setting the radius of the spherical area and the maximum number of sampling points in the spherical area, performing a spherical query operation, and obtaining a local point set.

[0009] S4, input the local point set into the joint recognition network for target recognition, and the joint recognition network includes a point cloud feature extraction module, a three-dimensional convolution module and an RPN network connected in sequence.

[0010] S5, sets the total loss function of the joint recognition network and evaluates the target recognition results.

[0011] Furthermore, the classified point cloud set in step S1 is expressed as: ; ; in, Represents a collection of classified point clouds, Is a collection In space A point in represents the spatial coordinates of a point, Represents the total number of points, Indicates the category of the point, , and There are three categories.

[0012] Furthermore, step S2 specifically includes the following steps: S2.1, in the point cloud collection with categories Randomly select a point from the set , specify the total number of sampling points as .

[0013] S2.2, Calculation The remaining points in the set The minimum distance value of each point in is taken as Each remaining point in the set distance.

[0014] S2.3, select The distance set of the remaining points in The farthest point is put into the set Repeat S2.2~S2.3 until the total number of sampling points reaches ,get : ; in, Is a collection The key points in express To Collection The Euclidean distance of each point in is and Indicates the selection of maximum and minimum value operations;

[0015] By key points The set of It is expressed as: ; in, Indicates key points The spatial coordinates of .

[0016] Furthermore, step S3 specifically includes the following steps: S3.1, for the set Each key point in ,by The coordinate point of is the center of the sphere, and the distance from the non-key point cloud with bottom sediment classification labels to the coordinates of the center of the sphere is calculated.

[0017] S3.2, select a value not exceeding the sphere radius The closest front points as the points in the local point set, the ball query operation is expressed as: ; in, is a local point set, Indicates the number of selected points.

[0018] Furthermore, in step S4, point cloud feature extraction is performed in a point cloud feature extraction module, and the point cloud feature extraction module includes a first, a second, and a third feature extraction layer, and specifically includes the following steps: S4.1, the input of the joint recognition network is a local point set Point cloud data in , local point set Points in It is expressed as: ; in, Contains the original coordinates of the point And the category of corresponding points .

[0019] S4.2, Integrate key points into point cloud data , get the extended features of the point : .

[0020] S4.3, in the first feature extraction layer, Each point in Mapped to In the dimensional space, the output feature is , the dimension is ,by As input, it is aggregated by the normalization layer, nonlinear layer and maximum pooling layer in the first feature extraction layer to obtain the spherical local feature. : ; in, Represents a max pooling operation.

[0021] S4.5, and After splicing, the points are sequentially passed through the second feature extraction layer and the third feature extraction layer, and then through the maximum pooling layer to obtain a comprehensive feature of a spherical point set.

[0022] Furthermore, the following steps are performed in the three-dimensional convolution module in step S4: S4.6, yes Execute steps S4.1 to S4.5 for each spherical point set, and obtain The spherical comprehensive characteristics will The spherical comprehensive features are spliced ​​together to make the spliced The spherical comprehensive features are used as the input of the 3D convolution module.

[0023] S4.7, design a three-dimensional convolution module, including: an encoding path and a decoding path, the encoding path and the decoding path respectively include three three-dimensional convolution layers and three three-dimensional deconvolution layers; one three-dimensional convolution layer corresponds to one three-dimensional deconvolution layer; in the encoding path, each three-dimensional convolution layer includes: a three-dimensional convolution, a batch normalization layer and a non-linear layer, in the decoding path, each three-dimensional deconvolution includes: a three-dimensional deconvolution, a batch normalization layer and a non-linear layer, in each three-dimensional deconvolution layer in the decoding path, the feature map obtained by the current layer is spliced ​​with the feature map of the corresponding level in the encoding path, and then a three-dimensional convolution operation is performed to fuse the information of the final encoding path and decoding path.

[0024] Furthermore, the following steps are performed in the RPN network in step S4: S4.8, the output of the 3D convolution module is used as the input of the RPN network.

[0025] S4.9, build the RPN network, including the bounding box detection head and the classification detection head.

[0026] S4.10, output features of the bounding box detection head It is expressed as: ; Among them, each dimension feature By the bounding box center coordinates , the length of the bounding box ,Width ,high and orientation angle express, is the maximum number of predictions.

[0027] Output features of classification detection head It is expressed as: ; in, Indicates the number of categories for classification, is the prediction result belonging to the positive category.

[0028] S4.11, set the bounding box feature of the true label to , set the category feature of the true label to .

[0029] Furthermore, step S5 specifically includes the following steps: S5.1, total loss function It is expressed as: ; in, , is the weight, is the regression loss, is the classification loss.

[0030] S5.2, regression loss is used to measure the error of the joint recognition network in predicting the target bounding box position, using and The KL divergence between the Laplace distributions of is used to measure the position error: ; in, is the center position coordinate loss, is the regression box size loss, is the orientation angle loss, is the vertex loss, , and They are the center position coordinates, regression box size, and orientation angle of the predicted value. , and They are the center position coordinates, regression box size, and orientation angle of the true value respectively; is the weight; represents the Laplace distribution.

[0031] S5.3, the classification loss is responsible for measuring the error of the joint recognition network in predicting the target classification label, using the focal loss function: ; in, is the regulating factor, is a logarithmic function, It is The category features of the true labels.

[0032] The present invention has the following beneficial effects: The present invention uses the original point cloud and the bottom classification result to make a comprehensive judgment, and improves the recognition accuracy. The original point cloud and the bottom classification result are jointly identified according to the recognition process of the present invention. First, the point cloud is corresponded point by point with the bottom classification result, and the point cloud with the bottom classification label is obtained, and then the point cloud is sampled at the farthest point to obtain the key point of downsampling, and the key point is the center of the sphere, and the points around the key point are divided into a local spherical point set using a sphere query. The divided point cloud is used as the input of the joint recognition network, and the joint recognition network is used to obtain an accurate recognition result, and the point cloud target recognition result is finally output. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the specific implementation or the prior art description. Obviously, the drawings described below are some implementations of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings: Figure 1 The structural diagram of the joint recognition network of the present invention is shown.

[0034] Figure 2 A point cloud target recognition result obtained by using the method provided by the present invention is shown.

[0035] Figure 3 Another point cloud target recognition result obtained by using the method provided by the present invention is shown. DETAILED DESCRIPTION

[0036] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0037] A point cloud target detection method based on laser radar fusion bottom information includes the following steps: S1, correspond the bottom classification result to each point in the point cloud, so that each point belongs to a category result, and obtain a point cloud set with classification.

[0038] S2, sampling the farthest point of the classified point cloud obtained in step S1 to obtain key points.

[0039] S3, taking the key point as the sphere center, setting the radius of the spherical area and the maximum number of sampling points in the spherical area, performing a spherical query operation, and obtaining a local point set.

[0040] S4, the local point set is input into the joint recognition network for target recognition, and the joint recognition network includes a point cloud feature extraction module, a three-dimensional convolution module and an RPN network connected in sequence. The RPN network is a region generation network.

[0041] S5, sets the total loss function of the joint recognition network and evaluates the target recognition results.

[0042] Specifically, taking the bottom sediment as an example, the bottom sediment is the material located at the bottom of the water, including rocks, seabed sediment, sand and mud, etc. Effective classification of the sediment can accurately divide different natural sediments such as sand and reefs and non-natural sediments such as metals, assist in the three-dimensional modeling of the underwater terrain, and facilitate the identification of targets in the underwater environment. The full waveform data provided by the lidar can reflect the characteristics of the bottom sediment in a variety of ways. By studying the data characteristics such as reflection characteristics and waveform characteristics, the type and composition of the sediment can be inferred. The sediment classification results are matched with each point in the point cloud so that each point belongs to a category result.

[0043] The classified point cloud set in step S1 is expressed as: ; ; in, Represents a collection of classified point clouds, Is a collection In space A point in represents the spatial coordinates of a point, Represents the total number of points, Indicates the category of the point, , and There are three categories, representing natural substrates such as metal, non-metal and mud and sand.

[0044] Specifically, the farthest-point-sampling (FPS) operation is performed on the classified point cloud, the total number of sampling points is set, and the sampling points obtained by the farthest-point sampling operation are determined as key points.

[0045] Step S2 specifically includes the following steps: S2.1, in the point cloud collection with categories Randomly select a point from the set , specify the total number of sampling points as .

[0046] S2.2, Calculation The remaining points in the set The minimum distance value of each point in is taken as Each remaining point in the set distance.

[0047] S2.3, select The distance set of the remaining points in The farthest point is put into the set Repeat S2.2~S2.3 until the total number of sampling points reaches ,get : ; in, Is a collection The key points in express To Collection The Euclidean distance of each point in is and Indicates the selection of maximum and minimum value operations;

[0048] By key points The set of It is expressed as: ; in, Indicates key points The spatial coordinates of .

[0049] Specifically, after completing the farthest point sampling, the key point is obtained. With the key point as the center of the sphere, the radius of the spherical area and the maximum number of sampling points in the spherical area are set, and the sphere query operation is performed. The number of spherical area point sets and the number of key points obtained by the sphere query operation are same.

[0050] Step S3 specifically includes the following steps: S3.1, for the set Each key point in ,by The coordinate point of is the center of the sphere, and the distance from the non-key point cloud with bottom sediment classification labels to the coordinates of the center of the sphere is calculated.

[0051] S3.2, select a value not exceeding the sphere radius The closest front points as the points in the local point set, the ball query operation is expressed as: ; in, is a local point set, Indicates the number of selected points.

[0052] When performing a ball query operation, there are two situations in which the number of point clouds within the ball radius exists: (1) When the number of point clouds within the ball radius is greater than or equal to the number of selected points , sort the distances from the key points at the center of the sphere, and select the front with the smallest distance points as local point sets. (2) When the number of point clouds within the sphere radius is less than , the point cloud quantity requirement cannot be met. In this case, the coordinates of the points that meet the requirement will be copied in order from near to far from the center of the sphere as missing points.

[0053] Specifically, Figure 1 As shown, in step S4, point cloud feature extraction is performed in a point cloud feature extraction module, and the point cloud feature extraction module includes a first, a second, and a third feature extraction layer, and specifically includes the following steps: S4.1, the input of the joint recognition network is a local point set Point cloud data in , local point set Points in It is expressed as: ; in, Contains the original coordinates of the point And the category of corresponding points .

[0054] S4.2, Integrate key points into point cloud data , which are points sorted by distance within a certain radius. By integrating the key point coordinates into these points, the expanded features of the points can be obtained. : .

[0055] S4.3, in the first feature extraction layer, Each point in Mapped to In the dimensional space, the output feature is , the dimension is By mapping the data into a high-dimensional space, the coordinates of the points and the bottom features are integrated, and the subsequent network can capture and learn complex and diverse data features in a higher-dimensional space. As input, it is aggregated by the normalization layer, nonlinear layer and maximum pooling layer in the first feature extraction layer to obtain the spherical local feature. : ; in, Represents a max pooling operation.

[0056] S4.5, and After splicing, each point contains its own high-dimensional features and spherical local features. After splicing, it passes through the second feature extraction layer and the third feature extraction layer in turn, and then passes through the maximum pooling layer to obtain a comprehensive feature of the spherical point set. After the data passes through multiple feature extraction layers, a more abstract and advanced representation can be learned, and feature representations from simple to complex, from local to global can be extracted. Finally, the features of all local spherical points are pooled to obtain a comprehensive feature of the spherical point set.

[0057] Specifically, the following steps are performed in the three-dimensional convolution module in step S4: S4.6, yes Execute steps S4.1 to S4.5 for each spherical point set, and obtain The spherical comprehensive characteristics will The spherical comprehensive features are spliced ​​together to make the spliced The spherical comprehensive features are used as the input of the 3D convolution module.

[0058] S4.7, design a three-dimensional convolution module, including: an encoding path and a decoding path, the encoding path and the decoding path respectively include three three-dimensional convolution layers and three three-dimensional deconvolution layers; one three-dimensional convolution layer corresponds to one three-dimensional deconvolution layer; in the encoding path, each three-dimensional convolution layer includes: a three-dimensional convolution, a batch normalization layer and a non-linear layer, in the decoding path, each three-dimensional deconvolution includes: a three-dimensional deconvolution, a batch normalization layer and a non-linear layer, in each three-dimensional deconvolution layer in the decoding path, the feature map obtained by the current layer is spliced ​​with the feature map of the corresponding level in the encoding path, and then a three-dimensional convolution operation is performed to fuse the information of the final encoding path and decoding path.

[0059] The splicing process uses a jump connection method similar to Unet, the purpose of which is to pass the information in the encoding path to the decoder path, make up for the information lost in the downsampling process, and improve the detection and recognition accuracy of the subsequent RPN network.

[0060] In the encoding path, the data goes through a series of 3D convolution, batch normalization, and nonlinear activation layers, and the size of the feature map gradually decreases while the number of channels gradually increases. This process extracts multi-scale features of the input data. The decoding path performs upsampling and restores the size of the feature map layer by layer. After each layer of deconvolution, the size of the feature map increases and the number of channels decreases.

[0061] At each layer of the decoding path, the feature map obtained by deconvolution of the current layer is concatenated with the feature map of the corresponding level in the encoding path. The corresponding level refers to feature maps with the same spatial size. The concatenation operation is performed on the channel dimension, without changing the depth, height and width of the feature map, but increasing the number of channels of the feature map.

[0062] Specifically, the following steps are performed in the RPN network in step S4: S4.8, the output of the 3D convolution module is used as the input of the RPN network.

[0063] S4.9, build the RPN network, including the bounding box detection head and the classification detection head.

[0064] S4.10, output features of the bounding box detection head It is expressed as: ; Among them, each dimension feature By the bounding box center coordinates , the length of the bounding box ,Width ,high and orientation angle express, is the maximum number of predictions.

[0065] Output features of classification detection head It is expressed as: ; in, Indicates the number of categories to be classified. If there is one more category, there is no category. is the prediction result belonging to the positive category.

[0066] S4.11, set the bounding box feature of the true label to , set the category feature of the true label to .

[0067] Specifically, step S5 includes the following steps: S5.1, in the target detection task, the overall loss of the network is set to regression loss and classification loss The total loss function is the sum of the two loss functions. It is expressed as: ; in, , is the weight, is the regression loss, is the classification loss.

[0068] S5.2, regression loss is used to measure the error of the joint recognition network in predicting the target bounding box position, using and The KL divergence between the Laplace distributions is used to measure the position error, since the predicted value and the true value There may be outliers between and The Kullback-Leibler (KL) divergence between the Laplace distributions of is used to measure the position error: ; in, is the center position coordinate loss, is the regression box size loss, is the orientation angle loss, is the vertex loss, , and They are the center position coordinates, regression box size, and orientation angle of the predicted value. , and They are the center position coordinates, regression box size, and orientation angle of the true value respectively; is the weight; represents the Laplace distribution.

[0069] In this regression loss, the center position coordinates, regression box size, and orientation angle all have independent losses. At the same time, a vertex loss is added , which is used to measure the distance between the eight vertices of the predicted box and the real box. Since the vertex position depends on the central coordinates, size and orientation angle of the regression box, the vertex loss can regularize the training of these parameters.

[0070] S5.3, the classification loss is responsible for measuring the error of the joint recognition network in predicting the target classification label, using the focal loss function: ; in, is the regulating factor, is a logarithmic function, It is The category features of the true labels.

[0071] Taking the point cloud data integrated with the bottom information as input, after FPS sampling, ball query grouping, joint network recognition and detection, the bounding box of the target and the detection results of the target category are obtained. The recognition results are as follows Figure 2 and Figure 3As shown in the figure above, the target objects placed in the pool are detected by generating point clouds and fusing the bottom information into the joint recognition network to obtain the detection results. There are metal cuboids, caissons and other objects placed in the pool. The metal cuboids are marked with green boxes in the figure, and the caissons are marked with blue boxes.

[0072] Based on the point cloud data, the present invention integrates the bottom data information to assist in detecting the point cloud target. The present invention uses sphere grouping as the point cloud encoding method, extracts the point cloud high-dimensional feature information through sphere grouping in the joint recognition stage, and integrates the spherical local and global feature information after information extraction.

[0073] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A point cloud target detection method based on laser radar fusion bottom information, characterized in that: The specific steps include: S1, correspond the bottom classification result to each point in the point cloud, so that each point belongs to a category result, and obtain a point cloud set with classification; S2, sampling the farthest point of the classified point cloud obtained in step S1 to obtain key points; S3, taking the key point as the sphere center, setting the radius of the spherical area and the maximum number of sampling points in the spherical area, performing a spherical query operation, and obtaining a local point set; S4, inputting the local point set into the joint recognition network for target recognition, the joint recognition network includes a point cloud feature extraction module, a three-dimensional convolution module and an RPN network connected in sequence; S5, sets the total loss function of the joint recognition network and evaluates the target recognition results.

2. The point cloud target detection method based on laser radar fusion bottom information according to claim 1 is characterized in that: The classified point cloud set in step S1 is expressed as: ； ； in, Represents a collection of classified point clouds, Is a collection In space A point in represents the spatial coordinates of a point, Represents the total number of points, Indicates the category of the point, , and There are three categories.

3. The point cloud target detection method based on laser radar fusion bottom information according to claim 1 is characterized in that: Step S2 specifically includes the following steps: S2.1, in the point cloud collection with categories Randomly select a point from the set , specify the total number of sampling points as ; S2.2, Calculation The remaining points in the set The minimum distance value of each point in is taken as Each remaining point in the set distance; S2.3, select The distance set of the remaining points in The farthest point is put into the set Repeat S2.2~S2.3 until the total number of sampling points reaches ,get : ； in, Is a collection The key points in express To Collection The Euclidean distance of each point in is and Indicates the selection of maximum and minimum value operations; By key points The set of It is expressed as: ； in, Indicates key points The spatial coordinates of .

4. The point cloud target detection method based on laser radar fusion bottom information according to claim 1 is characterized in that: Step S3 specifically includes the following steps: S3.1, for the set Each key point in ,by The coordinate point is taken as the center of the sphere, and the distance from the non-key point cloud with the bottom classification label to the coordinate of the center of the sphere is calculated; S3.2, select a value not exceeding the sphere radius The closest front points as the points in the local point set, the ball query operation is expressed as: ； in, is a local point set, Indicates the number of selected points.

5. The point cloud target detection method based on laser radar fusion bottom information according to claim 1 is characterized in that: In step S4, point cloud feature extraction is performed in a point cloud feature extraction module, and the point cloud feature extraction module includes a first, a second, and a third feature extraction layer, and specifically includes the following steps: S4.1, the input of the joint recognition network is a local point set Point cloud data in , local point set Points in It is expressed as: ； in, Contains the original coordinates of the point And the category of corresponding points ; S4.2, Integrate key points into point cloud data , get the extended features of the point : ； S4.3, in the first feature extraction layer, Each point in Mapped to In the dimensional space, the output feature is , the dimension is ,by As input, it is aggregated by the normalization layer, nonlinear layer and maximum pooling layer in the first feature extraction layer to obtain the spherical local feature. : ； in, Represents the maximum pooling operation; S4.5, and After splicing, the points are sequentially passed through the second feature extraction layer and the third feature extraction layer, and then through the maximum pooling layer to obtain a comprehensive feature of the spherical point set.

6. The point cloud target detection method based on laser radar fusion bottom information according to claim 5 is characterized in that: The following steps are performed in the three-dimensional convolution module in step S4: S4.6, yes Execute steps S4.1 to S4.5 for each spherical point set, and obtain The spherical comprehensive characteristics will The spherical comprehensive features are spliced ​​together to make the spliced The spherical comprehensive features are used as the input of the 3D convolution module; S4.7, design a three-dimensional convolution module, including: an encoding path and a decoding path, the encoding path and the decoding path respectively include three three-dimensional convolution layers and three three-dimensional deconvolution layers; one three-dimensional convolution layer corresponds to one three-dimensional deconvolution layer; in the encoding path, each three-dimensional convolution layer includes: a three-dimensional convolution, a batch normalization layer and a non-linear layer, in the decoding path, each three-dimensional deconvolution includes: a three-dimensional deconvolution, a batch normalization layer and a non-linear layer, in each three-dimensional deconvolution layer in the decoding path, the feature map obtained by the current layer is spliced ​​with the feature map of the corresponding level in the encoding path, and then a three-dimensional convolution operation is performed to fuse the information of the final encoding path and decoding path.

7. The point cloud target detection method based on laser radar fusion bottom information according to claim 6 is characterized in that: In step S4, the following steps are performed in the RPN network: S4.8, the output of the three-dimensional convolution module is used as the input of the RPN network; S4.9, construct the RPN network, including the bounding box detection head and the classification detection head; S4.10, output features of the bounding box detection head It is expressed as: ； Among them, each dimension feature By the bounding box center coordinates , the length of the bounding box ,Width ,high and orientation angle express, is the maximum number of predictions; Output features of classification detection head It is expressed as: ； in, Indicates the number of categories for classification, is the prediction result belonging to the positive category; S4.11, set the bounding box feature of the true label to , set the category feature of the true label to .

8. The point cloud target detection method based on laser radar fusion bottom information according to claim 1 is characterized in that: Step S5 specifically includes the following steps: S5.1, total loss function It is expressed as: ； in, , is the weight, is the regression loss, is the classification loss; S5.2, regression loss is used to measure the error of the joint recognition network in predicting the target bounding box position, using and The KL divergence between the Laplace distributions of is used to measure the position error: ； in, is the center position coordinate loss, is the regression box size loss, is the orientation angle loss, is the vertex loss, , and They are the center position coordinates, regression box size, and orientation angle of the predicted value. , and They are the center position coordinates of the true value, the regression box size, and the orientation angle; is the weight; represents the Laplace distribution; S5.3, the classification loss is responsible for measuring the error of the joint recognition network in predicting the target classification label, using the focal loss function: ； in, is the regulating factor, is a logarithmic function, It is The category features of the true labels.

Images